close to working

wip
2026-01-05 14:11:04 -05:00 · 2025-12-08 18:17:56 -08:00 · 2025-12-08 15:20:52 -08:00 · 2025-12-08 14:42:18 -08:00 · 2025-12-08 13:23:11 -08:00 · 2025-12-08 00:35:49 -08:00
21 changed files with 2061 additions and 728 deletions
--- a/.gitattributes
+++ b/.gitattributes
@@ -19,8 +19,6 @@ ml/backend/**/*.comp linguist-vendored
 ml/backend/**/*.glsl linguist-vendored
 ml/backend/**/CMakeLists.txt linguist-vendored

-app/webview linguist-vendored
-
 llama/build-info.cpp linguist-generated
 ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-embed.s linguist-generated

--- a/api/client.go
+++ b/api/client.go
@@ -226,14 +226,7 @@ func (c *Client) stream(ctx context.Context, method, path string, data any, fn f

 		bts := scanner.Bytes()
 		if err := json.Unmarshal(bts, &errorResponse); err != nil {
-			if response.StatusCode >= http.StatusBadRequest {
-				return StatusError{
-					StatusCode:   response.StatusCode,
-					Status:       response.Status,
-					ErrorMessage: string(bts),
-				}
-			}
-			return errors.New(string(bts))
+			return fmt.Errorf("unmarshal: %w", err)
 		}

 		if response.StatusCode == http.StatusUnauthorized {
--- a/api/client_test.go
+++ b/api/client_test.go
@@ -55,7 +55,6 @@ func TestClientFromEnvironment(t *testing.T) {
 type testError struct {
 	message    string
 	statusCode int
-	raw        bool // if true, write message as-is instead of JSON encoding
 }

 func (e testError) Error() string {
@@ -112,20 +111,6 @@ func TestClientStream(t *testing.T) {
 				},
 			},
 		},
-		{
-			name: "plain text error response",
-			responses: []any{
-				"internal server error",
-			},
-			wantErr: "internal server error",
-		},
-		{
-			name: "HTML error page",
-			responses: []any{
-				"<html><body>404 Not Found</body></html>",
-			},
-			wantErr: "404 Not Found",
-		},
 	}

 	for _, tc := range testCases {
@@ -150,12 +135,6 @@ func TestClientStream(t *testing.T) {
 						return
 					}

-					if str, ok := resp.(string); ok {
-						fmt.Fprintln(w, str)
-						flusher.Flush()
-						continue
-					}
-
 					if err := json.NewEncoder(w).Encode(resp); err != nil {
 						t.Fatalf("failed to encode response: %v", err)
 					}
@@ -194,10 +173,9 @@ func TestClientStream(t *testing.T) {

 func TestClientDo(t *testing.T) {
 	testCases := []struct {
-		name           string
-		response       any
-		wantErr        string
-		wantStatusCode int
+		name     string
+		response any
+		wantErr  string
 	}{
 		{
 			name: "immediate error response",
@@ -205,8 +183,7 @@ func TestClientDo(t *testing.T) {
 				message:    "test error message",
 				statusCode: http.StatusBadRequest,
 			},
-			wantErr:        "test error message",
-			wantStatusCode: http.StatusBadRequest,
+			wantErr: "test error message",
 		},
 		{
 			name: "server error response",
@@ -214,8 +191,7 @@ func TestClientDo(t *testing.T) {
 				message:    "internal error",
 				statusCode: http.StatusInternalServerError,
 			},
-			wantErr:        "internal error",
-			wantStatusCode: http.StatusInternalServerError,
+			wantErr: "internal error",
 		},
 		{
 			name: "successful response",
@@ -227,26 +203,6 @@ func TestClientDo(t *testing.T) {
 				Success: true,
 			},
 		},
-		{
-			name: "plain text error response",
-			response: testError{
-				message:    "internal server error",
-				statusCode: http.StatusInternalServerError,
-				raw:        true,
-			},
-			wantErr:        "internal server error",
-			wantStatusCode: http.StatusInternalServerError,
-		},
-		{
-			name: "HTML error page",
-			response: testError{
-				message:    "<html><body>404 Not Found</body></html>",
-				statusCode: http.StatusNotFound,
-				raw:        true,
-			},
-			wantErr:        "<html><body>404 Not Found</body></html>",
-			wantStatusCode: http.StatusNotFound,
-		},
 	}

 	for _, tc := range testCases {
@@ -254,16 +210,11 @@ func TestClientDo(t *testing.T) {
 			ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
 				if errResp, ok := tc.response.(testError); ok {
 					w.WriteHeader(errResp.statusCode)
-					if !errResp.raw {
-						err := json.NewEncoder(w).Encode(map[string]string{
-							"error": errResp.message,
-						})
-						if err != nil {
-							t.Fatal("failed to encode error response:", err)
-						}
-					} else {
-						// Write raw message (simulates non-JSON error responses)
-						fmt.Fprint(w, errResp.message)
+					err := json.NewEncoder(w).Encode(map[string]string{
+						"error": errResp.message,
+					})
+					if err != nil {
+						t.Fatal("failed to encode error response:", err)
 					}
 					return
 				}
@@ -290,15 +241,6 @@ func TestClientDo(t *testing.T) {
 				if err.Error() != tc.wantErr {
 					t.Errorf("error message mismatch: got %q, want %q", err.Error(), tc.wantErr)
 				}
-				if tc.wantStatusCode != 0 {
-					if statusErr, ok := err.(StatusError); ok {
-						if statusErr.StatusCode != tc.wantStatusCode {
-							t.Errorf("status code mismatch: got %d, want %d", statusErr.StatusCode, tc.wantStatusCode)
-						}
-					} else {
-						t.Errorf("expected StatusError, got %T", err)
-					}
-				}
 				return
 			}

--- a/cmd/chat_template/chat_template.py
+++ b/cmd/chat_template/chat_template.py
@@ -1,625 +0,0 @@
-#!/usr/bin/env python3
-# /// script
-# requires-python = ">=3.11"
-# dependencies = [
-#   "transformers>=4.57.0",
-#   "jinja2",
-#   "fastapi",
-#   "uvicorn",
-#   "pydantic",
-#   "requests",
-# ]
-# ///
-"""
-Chat Template Testing Tool
-
-Test HuggingFace chat templates against Ollama renderers.
-
-Usage:
-    # Run predefined test cases against a HuggingFace model
-    uv run cmd/chat_template/chat_template.py --model PrimeIntellect/INTELLECT-3
-
-    # Compare HuggingFace output with Ollama renderer
-    uv run cmd/chat_template/chat_template.py --model PrimeIntellect/INTELLECT-3 --ollama-model intellect3
-
-    # Start server for manual curl testing
-    uv run cmd/chat_template/chat_template.py --serve
-
-    # Show chat template for a model
-    uv run cmd/chat_template/chat_template.py --model PrimeIntellect/INTELLECT-3 --show-template
-"""
-
-import argparse
-import json
-import sys
-from typing import Any
-
-from transformers import AutoTokenizer
-
-
-TEST_CASES = [
-    {
-        "name": "basic_user_message",
-        "messages": [{"role": "user", "content": "Hello!"}],
-        "tools": None,
-    },
-    {
-        "name": "with_system_message",
-        "messages": [
-            {"role": "system", "content": "You are a helpful assistant."},
-            {"role": "user", "content": "Hello!"},
-        ],
-        "tools": None,
-    },
-    {
-        "name": "multi_turn_conversation",
-        "messages": [
-            {"role": "user", "content": "Hello"},
-            {"role": "assistant", "content": "Hi there!"},
-            {"role": "user", "content": "How are you?"},
-        ],
-        "tools": None,
-    },
-    {
-        "name": "with_tools",
-        "messages": [
-            {"role": "system", "content": "You are a helpful assistant."},
-            {"role": "user", "content": "What is the weather?"},
-        ],
-        "tools": [
-            {
-                "type": "function",
-                "function": {
-                    "name": "get_weather",
-                    "description": "Get the current weather",
-                    "parameters": {
-                        "type": "object",
-                        "required": ["location"],
-                        "properties": {
-                            "location": {"type": "string", "description": "The city"}
-                        },
-                    },
-                },
-            }
-        ],
-    },
-    {
-        "name": "tool_call_and_response",
-        "messages": [
-            {"role": "user", "content": "What is the weather in SF?"},
-            {
-                "role": "assistant",
-                "content": "Let me check the weather.",
-                "tool_calls": [
-                    {
-                        "id": "call_1",
-                        "type": "function",
-                        "function": {
-                            "name": "get_weather",
-                            "arguments": {"location": "San Francisco"},
-                        },
-                    }
-                ],
-            },
-            {"role": "tool", "content": '{"temperature": 68}', "tool_call_id": "call_1"},
-        ],
-        "tools": [
-            {
-                "type": "function",
-                "function": {
-                    "name": "get_weather",
-                    "description": "Get the current weather",
-                    "parameters": {
-                        "type": "object",
-                        "required": ["location"],
-                        "properties": {
-                            "location": {"type": "string", "description": "The city"}
-                        },
-                    },
-                },
-            }
-        ],
-    },
-    {
-        "name": "parallel_tool_calls",
-        "messages": [
-            {"role": "user", "content": "Get weather in SF and NYC"},
-            {
-                "role": "assistant",
-                "tool_calls": [
-                    {
-                        "id": "call_1",
-                        "type": "function",
-                        "function": {
-                            "name": "get_weather",
-                            "arguments": {"location": "San Francisco"},
-                        },
-                    },
-                    {
-                        "id": "call_2",
-                        "type": "function",
-                        "function": {
-                            "name": "get_weather",
-                            "arguments": {"location": "New York"},
-                        },
-                    },
-                ],
-            },
-            {"role": "tool", "content": '{"temperature": 68}', "tool_call_id": "call_1"},
-            {"role": "tool", "content": '{"temperature": 55}', "tool_call_id": "call_2"},
-        ],
-        "tools": [
-            {
-                "type": "function",
-                "function": {
-                    "name": "get_weather",
-                    "parameters": {
-                        "type": "object",
-                        "properties": {"location": {"type": "string"}},
-                    },
-                },
-            }
-        ],
-    },
-    # Thinking tests
-    {
-        "name": "assistant_with_thinking",
-        "messages": [
-            {"role": "user", "content": "What is 2+2?"},
-            {
-                "role": "assistant",
-                "content": "The answer is 4.",
-                "thinking": "Let me calculate: 2 + 2 = 4. This is basic arithmetic.",
-            },
-            {"role": "user", "content": "And 3+3?"},
-        ],
-        "tools": None,
-    },
-    {
-        "name": "thinking_with_tool_call",
-        "messages": [
-            {"role": "user", "content": "What's the weather in Paris?"},
-            {
-                "role": "assistant",
-                "content": "I'll check the weather for you.",
-                "thinking": "The user wants to know the weather in Paris. I should call the get_weather function.",
-                "tool_calls": [
-                    {
-                        "id": "call_1",
-                        "type": "function",
-                        "function": {
-                            "name": "get_weather",
-                            "arguments": {"location": "Paris"},
-                        },
-                    }
-                ],
-            },
-            {"role": "tool", "content": '{"temperature": 18, "condition": "cloudy"}', "tool_call_id": "call_1"},
-        ],
-        "tools": [
-            {
-                "type": "function",
-                "function": {
-                    "name": "get_weather",
-                    "description": "Get current weather",
-                    "parameters": {
-                        "type": "object",
-                        "properties": {"location": {"type": "string"}},
-                    },
-                },
-            }
-        ],
-    },
-    {
-        "name": "thinking_only_no_content",
-        "messages": [
-            {"role": "user", "content": "Think about this silently."},
-            {
-                "role": "assistant",
-                "content": "",  # HuggingFace requires content field
-                "thinking": "I'm thinking about this but won't respond with visible content.",
-            },
-            {"role": "user", "content": "What did you think?"},
-        ],
-        "tools": None,
-    },
-]
-
-# Cache for tokenizers
-_tokenizer_cache: dict[str, Any] = {}
-
-
-def get_tokenizer(model_name: str):
-    """Get or create tokenizer for the given model."""
-    if model_name not in _tokenizer_cache:
-        print(f"Loading tokenizer for {model_name}...", file=sys.stderr)
-        _tokenizer_cache[model_name] = AutoTokenizer.from_pretrained(model_name)
-    return _tokenizer_cache[model_name]
-
-
-def apply_template(
-    model: str,
-    messages: list[dict],
-    tools: list[dict] | None = None,
-) -> str:
-    """Apply HuggingFace chat template to messages."""
-    tokenizer = get_tokenizer(model)
-
-    if tools:
-        return tokenizer.apply_chat_template(
-            messages,
-            tools=tools,
-            tokenize=False,
-            add_generation_prompt=True,
-        )
-    else:
-        return tokenizer.apply_chat_template(
-            messages,
-            tokenize=False,
-            add_generation_prompt=True,
-        )
-
-
-def get_ollama_prompt(
-    ollama_model: str,
-    messages: list[dict],
-    tools: list[dict] | None = None,
-    ollama_host: str = "http://localhost:11434",
-) -> str | None:
-    """Get rendered prompt from Ollama using debug_render_only."""
-    import requests
-
-    # Convert messages to Ollama format
-    ollama_messages = []
-    for msg in messages:
-        ollama_msg = {"role": msg["role"]}
-        if "content" in msg:
-            ollama_msg["content"] = msg["content"]
-        if "thinking" in msg:
-            ollama_msg["thinking"] = msg["thinking"]
-        if "tool_calls" in msg:
-            # Convert tool_calls to Ollama format
-            tool_calls = []
-            for tc in msg["tool_calls"]:
-                tool_call = {
-                    "function": {
-                        "name": tc["function"]["name"],
-                        "arguments": tc["function"]["arguments"],
-                    }
-                }
-                if "id" in tc:
-                    tool_call["id"] = tc["id"]
-                tool_calls.append(tool_call)
-            ollama_msg["tool_calls"] = tool_calls
-        if "tool_call_id" in msg:
-            ollama_msg["tool_call_id"] = msg["tool_call_id"]
-        ollama_messages.append(ollama_msg)
-
-    payload = {
-        "model": ollama_model,
-        "messages": ollama_messages,
-        "stream": False,
-        "_debug_render_only": True,
-    }
-
-    if tools:
-        payload["tools"] = tools
-
-    try:
-        resp = requests.post(f"{ollama_host}/api/chat", json=payload, timeout=30)
-        resp.raise_for_status()
-        data = resp.json()
-        # Field name is _debug_info with underscore prefix
-        if "_debug_info" in data and "rendered_template" in data["_debug_info"]:
-            return data["_debug_info"]["rendered_template"]
-        return None
-    except requests.exceptions.ConnectionError:
-        print(f"  [ERROR] Cannot connect to Ollama at {ollama_host}", file=sys.stderr)
-        return None
-    except Exception as e:
-        print(f"  [ERROR] Ollama request failed: {e}", file=sys.stderr)
-        return None
-
-
-def compute_diff(hf_prompt: str, ollama_prompt: str) -> str:
-    """Compute a unified diff between HuggingFace and Ollama prompts."""
-    import difflib
-
-    hf_lines = hf_prompt.splitlines(keepends=True)
-    ollama_lines = ollama_prompt.splitlines(keepends=True)
-
-    diff = difflib.unified_diff(
-        ollama_lines,
-        hf_lines,
-        fromfile="Ollama",
-        tofile="HuggingFace",
-        lineterm="",
-    )
-    return "".join(diff)
-
-
-def print_test_output(
-    name: str,
-    messages: list[dict],
-    tools: list[dict] | None,
-    hf_prompt: str,
-    ollama_prompt: str | None = None,
-    as_repr: bool = False,
-):
-    """Print test output in a format suitable for Go test creation and LLM diffing."""
-    print(f"\n{'='*60}")
-    print(f"Test: {name}")
-    print("=" * 60)
-    print("\n--- Input Messages ---")
-    print(json.dumps(messages, indent=2))
-    if tools:
-        print("\n--- Tools ---")
-        print(json.dumps(tools, indent=2))
-
-    if ollama_prompt is not None:
-        # Comparison mode
-        if hf_prompt == ollama_prompt:
-            print("\n--- Result: MATCH ---")
-            print("\n--- Prompt (both identical) ---")
-            if as_repr:
-                print(repr(hf_prompt))
-            else:
-                print(hf_prompt)
-        else:
-            print("\n--- Result: MISMATCH ---")
-            print("\n--- HuggingFace Prompt ---")
-            if as_repr:
-                print(repr(hf_prompt))
-            else:
-                print(hf_prompt)
-            print("\n--- Ollama Prompt ---")
-            if as_repr:
-                print(repr(ollama_prompt))
-            else:
-                print(ollama_prompt)
-            print("\n--- Diff (Ollama -> HuggingFace) ---")
-            diff = compute_diff(hf_prompt, ollama_prompt)
-            if diff:
-                print(diff)
-            else:
-                print("(no line-level diff, check whitespace)")
-    else:
-        # HuggingFace only mode
-        print("\n--- HuggingFace Prompt ---")
-        if as_repr:
-            print(repr(hf_prompt))
-        else:
-            print(hf_prompt)
-
-    print("=" * 60)
-
-
-def run_tests(
-    model: str,
-    as_repr: bool = False,
-    test_filter: str | None = None,
-    ollama_model: str | None = None,
-    ollama_host: str = "http://localhost:11434",
-):
-    """Run all predefined test cases against a model."""
-    if ollama_model:
-        print(f"\nComparing HuggingFace ({model}) vs Ollama ({ollama_model})\n")
-    else:
-        print(f"\nRunning tests against: {model}\n")
-
-    matches = 0
-    mismatches = 0
-    errors = 0
-
-    for test_case in TEST_CASES:
-        name = test_case["name"]
-        messages = test_case["messages"]
-        tools = test_case["tools"]
-
-        # Filter tests if specified
-        if test_filter and test_filter.lower() not in name.lower():
-            continue
-
-        try:
-            hf_prompt = apply_template(model, messages, tools)
-
-            ollama_prompt = None
-            if ollama_model:
-                ollama_prompt = get_ollama_prompt(
-                    ollama_model, messages, tools, ollama_host
-                )
-                if ollama_prompt is None:
-                    errors += 1
-                elif hf_prompt == ollama_prompt:
-                    matches += 1
-                else:
-                    mismatches += 1
-
-            print_test_output(
-                name, messages, tools, hf_prompt, ollama_prompt, as_repr=as_repr
-            )
-        except Exception as e:
-            errors += 1
-            print(f"\n{'='*60}")
-            print(f"Test: {name} - FAILED")
-            print(f"--- Input Messages ---")
-            print(json.dumps(messages, indent=2))
-            if tools:
-                print(f"--- Tools ---")
-                print(json.dumps(tools, indent=2))
-            print(f"--- Error ---")
-            print(f"{e}")
-            print("=" * 60)
-
-    # Print summary if comparing
-    if ollama_model:
-        total = matches + mismatches + errors
-        print(f"\n{'='*60}")
-        print("SUMMARY")
-        print("=" * 60)
-        print(f"  Total:      {total}")
-        print(f"  Matches:    {matches}")
-        print(f"  Mismatches: {mismatches}")
-        print(f"  Errors:     {errors}")
-        print("=" * 60)
-
-
-def show_template(model: str):
-    """Show the chat template for a model."""
-    tokenizer = get_tokenizer(model)
-    print(f"\nChat template for {model}:\n")
-    print("-" * 60)
-    print(tokenizer.chat_template)
-    print("-" * 60)
-
-
-def start_server(host: str = "0.0.0.0", port: int = 8000):
-    """Start the FastAPI server for manual testing."""
-    from typing import Optional, List, Dict, Any as TypingAny
-
-    from fastapi import FastAPI, HTTPException
-    from pydantic import BaseModel
-    import uvicorn
-
-    class Message(BaseModel):
-        role: str
-        content: Optional[str] = None
-        tool_calls: Optional[List[Dict[str, TypingAny]]] = None
-        tool_call_id: Optional[str] = None
-
-    class GeneratePromptRequest(BaseModel):
-        messages: List[Message]
-        model: str = "PrimeIntellect/INTELLECT-3"
-        tools: Optional[List[Dict[str, TypingAny]]] = None
-        inject_tools_as_functions: bool = False
-
-    class GeneratePromptResponse(BaseModel):
-        prompt: str
-        model: str
-
-    app = FastAPI(title="HuggingFace Prompt Generator", version="1.0.0")
-
-    @app.post("/generate-prompt", response_model=GeneratePromptResponse)
-    async def generate_prompt(request: GeneratePromptRequest):
-        try:
-            messages = []
-            for msg in request.messages:
-                message_dict = {"role": msg.role}
-                if msg.content is not None:
-                    message_dict["content"] = msg.content
-                if msg.tool_calls is not None:
-                    tool_calls = []
-                    for tc in msg.tool_calls:
-                        tc_copy = tc.copy()
-                        if "function" in tc_copy and "arguments" in tc_copy["function"]:
-                            args = tc_copy["function"]["arguments"]
-                            if isinstance(args, str):
-                                try:
-                                    tc_copy["function"]["arguments"] = json.loads(args)
-                                except json.JSONDecodeError:
-                                    pass
-                        tool_calls.append(tc_copy)
-                    message_dict["tool_calls"] = tool_calls
-                if msg.tool_call_id is not None:
-                    message_dict["tool_call_id"] = msg.tool_call_id
-                messages.append(message_dict)
-
-            prompt = apply_template(request.model, messages, request.tools)
-            return GeneratePromptResponse(prompt=prompt, model=request.model)
-        except Exception as e:
-            raise HTTPException(status_code=500, detail=str(e))
-
-    @app.get("/health")
-    async def health_check():
-        return {"status": "healthy"}
-
-    print(f"Starting server on http://{host}:{port}")
-    print("Endpoints:")
-    print("  POST /generate-prompt - Generate prompt from messages")
-    print("  GET  /health          - Health check")
-    uvicorn.run(app, host=host, port=port)
-
-
-def main():
-    parser = argparse.ArgumentParser(
-        description="HuggingFace Prompt Testing Tool",
-        formatter_class=argparse.RawDescriptionHelpFormatter,
-        epilog=__doc__,
-    )
-    parser.add_argument(
-        "--model",
-        "-m",
-        type=str,
-        help="HuggingFace model name (e.g., PrimeIntellect/INTELLECT-3)",
-    )
-    parser.add_argument(
-        "--ollama-model",
-        "-o",
-        type=str,
-        help="Ollama model name to compare against (e.g., qwen3-coder)",
-    )
-    parser.add_argument(
-        "--ollama-host",
-        type=str,
-        default="http://localhost:11434",
-        help="Ollama server URL (default: http://localhost:11434)",
-    )
-    parser.add_argument(
-        "--serve",
-        "-s",
-        action="store_true",
-        help="Start FastAPI server for manual curl testing",
-    )
-    parser.add_argument(
-        "--port",
-        "-p",
-        type=int,
-        default=8000,
-        help="Server port (default: 8000)",
-    )
-    parser.add_argument(
-        "--show-template",
-        "-t",
-        action="store_true",
-        help="Show the chat template for the model",
-    )
-    parser.add_argument(
-        "--repr",
-        "-r",
-        action="store_true",
-        help="Output prompts as Python repr (shows escape sequences)",
-    )
-    parser.add_argument(
-        "--filter",
-        "-f",
-        type=str,
-        help="Filter tests by name (substring match)",
-    )
-
-    args = parser.parse_args()
-
-    if args.serve:
-        start_server(port=args.port)
-    elif args.model:
-        if args.show_template:
-            show_template(args.model)
-        else:
-            run_tests(
-                args.model,
-                as_repr=args.repr,
-                test_filter=args.filter,
-                ollama_model=args.ollama_model,
-                ollama_host=args.ollama_host,
-            )
-    else:
-        parser.print_help()
-        print("\nExample usage:")
-        print("  uv run cmd/chat_template/chat_template.py --model PrimeIntellect/INTELLECT-3")
-        print("  uv run cmd/chat_template/chat_template.py --model Qwen/Qwen3-Coder-480B-A35B-Instruct --ollama-model qwen3-coder")
-        print("  uv run cmd/chat_template/chat_template.py --serve")
-        sys.exit(1)
-
-
-if __name__ == "__main__":
-    main()
-
--- a/cmd/testolmo/main.go
+++ b/cmd/testolmo/main.go
@@ -0,0 +1,148 @@
+package main
+
+import (
+	"context"
+	"fmt"
+	"log"
+
+	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/ml"
+	"github.com/ollama/ollama/model"
+	"github.com/ollama/ollama/model/input"
+	_ "github.com/ollama/ollama/model/models" // Register all models
+	"github.com/ollama/ollama/model/renderers"
+	"github.com/ollama/ollama/sample"
+)
+
+func main() {
+	modelPath := "/Users/parth/.ollama/models/blobs/sha256-a87e10578f328b087f888ac7bd1018555e26028a1130980f20312b4de3a10d70"
+
+	fmt.Println("Loading OLMo model...")
+	m, err := model.New(modelPath, ml.BackendParams{AllocMemory: true})
+	if err != nil {
+		log.Fatal(err)
+	}
+
+	if err := m.Backend().Load(context.Background(), func(f float32) {}); err != nil {
+		log.Fatal(err)
+	}
+
+	fmt.Println("✅ Model loaded successfully!")
+
+	// Initialize the cache
+	cache := m.Config().Cache
+	if cache != nil {
+		// Initialize with reasonable defaults:
+		// - dtype: F16
+		// - maxSequences: 1 (single sequence)
+		// - capacity: 2048 (context length)
+		// - maxBatch: 512
+		cache.Init(m.Backend(), ml.DTypeF16, 1, 2048, 512)
+		fmt.Printf("✅ Cache initialized (type: %T)\n", cache)
+	}
+
+	// Use the olmo3 renderer to format the prompt properly
+	messages := []api.Message{
+		{Role: "user", Content: "wagwan"},
+	}
+	// prompt := "Question: What is machine learning? Answer:"
+	prompt, err := renderers.RenderWithRenderer("olmo3", messages, nil, nil)
+	if err != nil {
+		log.Fatal(err)
+	}
+	// prompt = prompt[:len(prompt)]
+	// prompt := "Question: What is machine learning? Answer:"
+	fmt.Printf("\nRendered prompt:\n%s\n", prompt)
+
+	tp := m.(model.TextProcessor)
+	tokens, err := tp.Encode(prompt, false)
+	if err != nil {
+		log.Fatal(err)
+	}
+
+	fmt.Printf("Tokens: %v (count: %d)\n", tokens, len(tokens))
+
+	// Generate 20 tokens
+	maxTokens := 20
+	generated := make([]int32, 0, maxTokens)
+
+	// Create sampler (temperature=0 for greedy sampling)
+	sampler := sample.NewSampler(0, 0, 0, 0, -1, nil)
+
+	for i := 0; i < maxTokens; i++ {
+		// Create a new context for each generation step to avoid memory buildup
+		ctx := m.Backend().NewContext()
+
+		var inputTokens []int32
+		var positions []int32
+
+		if i == 0 {
+			// First iteration: process all prompt tokens
+			inputTokens = tokens
+			positions = make([]int32, len(tokens))
+			for j := range positions {
+				positions[j] = int32(j)
+			}
+		} else {
+			// Subsequent iterations: only process the newly generated token
+			// The last token is at position len(tokens)-1 (its index in the sequence)
+			inputTokens = []int32{tokens[len(tokens)-1]}
+			positions = []int32{int32(len(tokens) - 1)}
+		}
+
+		sequences := make([]int, len(inputTokens))
+		// All tokens belong to sequence 0
+
+		inputsTensor := ctx.Input().FromInts(inputTokens, len(inputTokens))
+		outputs := ctx.Input().FromInts([]int32{int32(len(inputTokens) - 1)}, 1)
+
+		batch := input.Batch{
+			Inputs:    inputsTensor,
+			Positions: positions,
+			Sequences: sequences,
+			Outputs:   outputs,
+		}
+
+		// Forward pass (model.Forward handles cache.StartForward internally)
+		logits, err := model.Forward(ctx, m, batch)
+		if err != nil {
+			ctx.Close()
+			log.Fatal(err)
+		}
+
+		logits = logits.Contiguous(ctx)
+		ctx.Forward(logits).Compute(logits)
+
+		logitValues := logits.Floats()
+
+		// Sample next token
+		nextToken, err := sampler.Sample(logitValues)
+		if err != nil {
+			ctx.Close()
+			log.Fatal(err)
+		}
+
+		// Close context before moving to next iteration
+		ctx.Close()
+
+		generated = append(generated, nextToken)
+		tokens = append(tokens, nextToken)
+
+		// Decode and print
+		decoded, _ := tp.Decode([]int32{nextToken})
+		fmt.Print(decoded)
+
+		// Stop on EOS or <|im_end|>
+		if nextToken == 2 || nextToken == 1 { // Common EOS tokens
+			break
+		}
+		// Check if we generated <|im_end|> (stop token for chat)
+		if decoded == "<|im_end|>" {
+			break
+		}
+	}
+
+	fmt.Println("\n\n✅ Generation completed!")
+	fullText, _ := tp.Decode(generated)
+	fmt.Printf("Generated: %s\n", fullText)
+}
--- a/convert/convert.go
+++ b/convert/convert.go
@@ -200,6 +200,8 @@ func ConvertModel(fsys fs.FS, f *os.File) error {
 		conv = &qwen25VLModel{}
 	case "Qwen3VLForConditionalGeneration", "Qwen3VLMoeForConditionalGeneration":
 		conv = &qwen3VLModel{}
+	case "OLMo2ForCausalLM", "Olmo2ForCausalLM", "OLMo3ForCausalLM", "Olmo3ForCausalLM":
+		conv = &olmoModel{}
 	case "BertModel":
 		conv = &bertModel{}
 	case "CohereForCausalLM":
--- a/convert/convert_olmo.go
+++ b/convert/convert_olmo.go
@@ -0,0 +1,94 @@
+package convert
+
+import (
+	"cmp"
+
+	"github.com/ollama/ollama/fs/ggml"
+)
+
+type olmoModel struct {
+	ModelParameters
+
+	HiddenSize            uint32   `json:"hidden_size"`
+	NumHiddenLayers       uint32   `json:"num_hidden_layers"`
+	IntermediateSize      uint32   `json:"intermediate_size"`
+	NumAttentionHeads     uint32   `json:"num_attention_heads"`
+	NumKeyValueHeads      uint32   `json:"num_key_value_heads"`
+	MaxPositionEmbeddings uint32   `json:"max_position_embeddings"`
+	RMSNormEPS            float32  `json:"rms_norm_eps"`
+	RopeTheta             float32  `json:"rope_theta"`
+	ClampKQV              float32  `json:"f_clamp_kqv"`
+	SlidingWindow         uint32   `json:"sliding_window"`
+	LayerTypes            []string `json:"layer_types"`
+}
+
+var _ ModelConverter = (*olmoModel)(nil)
+
+func (p *olmoModel) KV(t *Tokenizer) ggml.KV {
+	kv := p.ModelParameters.KV(t)
+	kv["general.architecture"] = "olmo"
+	kv["olmo.block_count"] = p.NumHiddenLayers
+	kv["olmo.context_length"] = p.MaxPositionEmbeddings
+	kv["olmo.embedding_length"] = p.HiddenSize
+	kv["olmo.feed_forward_length"] = p.IntermediateSize
+	kv["olmo.attention.head_count"] = p.NumAttentionHeads
+	kv["olmo.attention.head_count_kv"] = cmp.Or(p.NumKeyValueHeads, p.NumAttentionHeads)
+
+	if p.RopeTheta > 0 {
+		kv["olmo.rope.freq_base"] = p.RopeTheta
+	} else {
+		kv["olmo.rope.freq_base"] = float32(10000.0)
+	}
+
+	if p.RMSNormEPS > 0 {
+		kv["olmo.attention.layer_norm_rms_epsilon"] = p.RMSNormEPS
+	}
+
+	if p.ClampKQV > 0 {
+		kv["olmo.attention.clamp_kqv"] = p.ClampKQV
+	}
+
+	if p.SlidingWindow > 0 {
+		kv["olmo.attention.sliding_window"] = p.SlidingWindow
+	}
+
+	if len(p.LayerTypes) > 0 {
+		kv["olmo.attention.layer_types"] = p.LayerTypes
+	}
+
+	return kv
+}
+
+func (p *olmoModel) Tensors(ts []Tensor) []*ggml.Tensor {
+	out := make([]*ggml.Tensor, 0, len(ts))
+	for _, t := range ts {
+		out = append(out, &ggml.Tensor{
+			Name:     t.Name(),
+			Kind:     t.Kind(),
+			Shape:    t.Shape(),
+			WriterTo: t,
+		})
+	}
+
+	return out
+}
+
+func (p *olmoModel) Replacements() []string {
+	return []string{
+		"lm_head", "output",
+		"model.embed_tokens", "token_embd",
+		"model.layers", "blk",
+		"model.norm", "output_norm",
+		"self_attn.q_proj", "attn_q",
+		"self_attn.k_proj", "attn_k",
+		"self_attn.v_proj", "attn_v",
+		"self_attn.o_proj", "attn_output",
+		"self_attn.q_norm", "attn_q_norm",
+		"self_attn.k_norm", "attn_k_norm",
+		"post_attention_layernorm", "post_attention_norm",
+		"post_feedforward_layernorm", "post_ffw_norm",
+		"mlp.gate_proj", "ffn_gate",
+		"mlp.down_proj", "ffn_down",
+		"mlp.up_proj", "ffn_up",
+	}
+}
--- a/discover/runner.go
+++ b/discover/runner.go
@@ -65,7 +65,6 @@ func GPUDevices(ctx context.Context, runners []ml.FilteredRunnerDiscovery) []ml.
 		}

 		slog.Info("discovering available GPUs...")
-		detectIncompatibleLibraries()

 		// Warn if any user-overrides are set which could lead to incorrect GPU discovery
 		overrideWarnings()
@@ -99,9 +98,6 @@ func GPUDevices(ctx context.Context, runners []ml.FilteredRunnerDiscovery) []ml.
 					continue
 				} else if jetpack != "" && filepath.Base(dir) != "cuda_"+jetpack {
 					continue
-				} else if jetpack == "" && strings.Contains(filepath.Base(dir), "cuda_jetpack") {
-					slog.Debug("jetpack not detected (set JETSON_JETPACK or OLLAMA_LLM_LIBRARY to override), skipping", "libDir", dir)
-					continue
 				} else if !envconfig.EnableVulkan() && strings.Contains(filepath.Base(dir), "vulkan") {
 					slog.Info("experimental Vulkan support disabled.  To enable, set OLLAMA_VULKAN=1")
 					continue
@@ -488,16 +484,3 @@ func overrideWarnings() {
 		slog.Warn("if GPUs are not correctly discovered, unset and try again")
 	}
 }
-
-func detectIncompatibleLibraries() {
-	if runtime.GOOS != "windows" {
-		return
-	}
-	basePath, err := exec.LookPath("ggml-base.dll")
-	if err != nil || basePath == "" {
-		return
-	}
-	if !strings.HasPrefix(basePath, ml.LibOllamaPath) {
-		slog.Warn("potentially incompatible library detected in PATH", "location", basePath)
-	}
-}
--- a/docs/faq.mdx
+++ b/docs/faq.mdx
@@ -57,13 +57,8 @@ ollama ps
 ```

 <Info>
-
-**Output**:
-
-```
-NAME        ID            SIZE    PROCESSOR   UNTIL
-llama3:70b  bcfb190ca3a7  42 GB   100% GPU    4 minutes from now
-```
+  **Output**: ``` NAME ID SIZE PROCESSOR UNTIL llama3:70b bcfb190ca3a7 42 GB
+  100% GPU 4 minutes from now ```
 </Info>

 The `Processor` column will show which memory the model was loaded in to:
@@ -390,4 +385,4 @@ Ollama for Windows and macOS register as a login item during installation.  You
 - In `Task Manager` go to the `Startup apps` tab, search for `ollama` then click `Disable`

 **MacOS**
- Open `Settings` and search for "Login Items", find the `Ollama` entry under "Allow in the Background`, then click the slider to disable.
+- Open `Settings` and search for "Login Items", find the `Ollama` entry under "Allow in the Background`, then click the slider to disable.
--- a/docs/modelfile.mdx
+++ b/docs/modelfile.mdx
@@ -149,6 +149,9 @@ PARAMETER <parameter> <parametervalue>

 | Parameter      | Description                                                                                                                                                                                                                                                                                                                                                                     | Value Type | Example Usage        |
 | -------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ---------- | -------------------- |
+| mirostat       | Enable Mirostat sampling for controlling perplexity. (default: 0, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0)                                                                                                                                                                                                                                                                 | int        | mirostat 0           |
+| mirostat_eta   | Influences how quickly the algorithm responds to feedback from the generated text. A lower learning rate will result in slower adjustments, while a higher learning rate will make the algorithm more responsive. (Default: 0.1)                                                                                                                                                | float      | mirostat_eta 0.1     |
+| mirostat_tau   | Controls the balance between coherence and diversity of the output. A lower value will result in more focused and coherent text. (Default: 5.0)                                                                                                                                                                                                                                 | float      | mirostat_tau 5.0     |
 | num_ctx        | Sets the size of the context window used to generate the next token. (Default: 2048)                                                                                                                                                                                                                                                                                            | int        | num_ctx 4096         |
 | repeat_last_n  | Sets how far back for the model to look back to prevent repetition. (Default: 64, 0 = disabled, -1 = num_ctx)                                                                                                                                                                                                                                                                   | int        | repeat_last_n 64     |
 | repeat_penalty | Sets how strongly to penalize repetitions. A higher value (e.g., 1.5) will penalize repetitions more strongly, while a lower value (e.g., 0.9) will be more lenient. (Default: 1.1)                                                                                                                                                                                             | float      | repeat_penalty 1.1   |
--- a/fs/ggml/ggml.go
+++ b/fs/ggml/ggml.go
@@ -252,6 +252,7 @@ func (kv KV) OllamaEngineRequired() bool {
 		"deepseekocr",
 		"deepseek2",
 		"nomic-bert",
+		"olmo2",
 	}, kv.Architecture())
 }

--- a/model/models/models.go
+++ b/model/models/models.go
@@ -13,6 +13,7 @@ import (
 	_ "github.com/ollama/ollama/model/models/mistral3"
 	_ "github.com/ollama/ollama/model/models/mllama"
 	_ "github.com/ollama/ollama/model/models/nomicbert"
+	_ "github.com/ollama/ollama/model/models/olmo"
 	_ "github.com/ollama/ollama/model/models/qwen2"
 	_ "github.com/ollama/ollama/model/models/qwen25vl"
 	_ "github.com/ollama/ollama/model/models/qwen3"
--- a/model/models/olmo/model.go
+++ b/model/models/olmo/model.go
@@ -0,0 +1,271 @@
+package olmo
+
+import (
+	"cmp"
+	"math"
+
+	"github.com/ollama/ollama/fs"
+	"github.com/ollama/ollama/kvcache"
+	"github.com/ollama/ollama/ml"
+	"github.com/ollama/ollama/ml/nn"
+	"github.com/ollama/ollama/ml/nn/fast"
+	"github.com/ollama/ollama/ml/nn/rope"
+	"github.com/ollama/ollama/model"
+	"github.com/ollama/ollama/model/input"
+)
+
+type Options struct {
+	hiddenSize, numHeads, numKVHeads int
+	headDim, ropeDim                 int
+	eps, ropeBase, ropeScale         float32
+	clampKQV                         float32
+
+	originalContextLength int
+	attnFactor            float32
+	slidingWindow         int32
+	slidingWindowPattern  []bool // per-layer SWA pattern (true = SWA, false = full attention)
+}
+
+type Model struct {
+	model.Base
+	model.TextProcessor
+
+	TokenEmbedding *nn.Embedding `gguf:"token_embd"`
+	Layers         []Layer       `gguf:"blk"`
+	OutputNorm     *nn.RMSNorm   `gguf:"output_norm"`
+	Output         *nn.Linear    `gguf:"output,alt:token_embd"`
+
+	Options
+}
+
+func New(c fs.Config) (model.Model, error) {
+	vocabulary := model.Vocabulary{
+		Values: c.Strings("tokenizer.ggml.tokens"),
+		Scores: c.Floats("tokenizer.ggml.scores"),
+		Types:  c.Ints("tokenizer.ggml.token_type"),
+		Merges: c.Strings("tokenizer.ggml.merges"),
+		AddBOS: c.Bool("tokenizer.ggml.add_bos_token", true),
+		BOS:    []int32{int32(c.Uint("tokenizer.ggml.bos_token_id"))},
+		AddEOS: c.Bool("tokenizer.ggml.add_eos_token", false),
+		EOS: append(
+			[]int32{int32(c.Uint("tokenizer.ggml.eos_token_id"))},
+			c.Ints("tokenizer.ggml.eos_token_ids")...,
+		),
+	}
+
+	if c.String("tokenizer.ggml.model") != "gpt2" {
+		return nil, model.ErrUnsupportedTokenizer
+	}
+
+	var pretokenizers []string
+	if c.String("tokenizer.ggml.pre") != "default" {
+		pretokenizers = []string{
+			"(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
+		}
+	}
+	processor := model.NewBytePairEncoding(&vocabulary, pretokenizers...)
+
+	slidingWindow := int32(c.Uint("attention.sliding_window"))
+	slidingWindowPattern := c.Bools("attention.sliding_window_pattern")
+
+	m := Model{
+		TextProcessor: processor,
+		Layers:        make([]Layer, c.Uint("block_count")),
+		Options: Options{
+			hiddenSize:            int(c.Uint("embedding_length")),
+			numHeads:              int(c.Uint("attention.head_count")),
+			numKVHeads:            int(c.Uint("attention.head_count_kv")),
+			headDim:               int(c.Uint("attention.key_length")),
+			ropeDim:               int(c.Uint("rope.dimension_count")),
+			eps:                   c.Float("attention.layer_norm_rms_epsilon"),
+			ropeBase:              c.Float("rope.freq_base", 1e4),
+			ropeScale:             c.Float("rope.scaling.factor", 1),
+			clampKQV:              c.Float("attention.clamp_kqv", 0),
+			originalContextLength: int(c.Uint("rope.scaling.original_context_length")),
+			attnFactor:            c.Float("rope.scaling.attn_factor", 1),
+			slidingWindow:         slidingWindow,
+			slidingWindowPattern:  slidingWindowPattern,
+		},
+	}
+
+	// OLMo3 uses interleaved sliding window attention (every 4th layer is full attention)
+	m.Cache = kvcache.NewWrapperCache(
+		kvcache.NewSWACache(slidingWindow, m.Shift),
+		kvcache.NewCausalCache(m.Shift),
+	)
+
+	return &m, nil
+}
+
+type SelfAttention struct {
+	Query       *nn.Linear  `gguf:"attn_q"`
+	Key         *nn.Linear  `gguf:"attn_k"`
+	Value       *nn.Linear  `gguf:"attn_v"`
+	Output      *nn.Linear  `gguf:"attn_output"`
+	QNorm       *nn.RMSNorm `gguf:"attn_q_norm"`
+	KNorm       *nn.RMSNorm `gguf:"attn_k_norm"`
+	RopeFactors ml.Tensor   `gguf:"rope_freqs.weight"`
+}
+
+func (o *Options) ropeOptions(factors ml.Tensor, isSWA bool) []func(*rope.Options) {
+	opts := []func(*rope.Options){
+		rope.WithFactors(factors),
+	}
+
+	if o.originalContextLength > 0 {
+		if isSWA {
+			// For SWA layers, use regular rope with no YaRN scaling
+			// ext_factor=0.0, attn_factor=1.0 per llama.cpp
+			opts = append(opts,
+				rope.WithOriginalContextLength(o.originalContextLength),
+				rope.WithExtrapolationFactor(0.),
+				rope.WithAttentionFactor(1.),
+			)
+		} else {
+			// For full attention layers, use YaRN scaling
+			opts = append(opts,
+				rope.WithOriginalContextLength(o.originalContextLength),
+				rope.WithExtrapolationFactor(1.),
+				rope.WithAttentionFactor(o.attnFactor),
+			)
+		}
+	}
+
+	return opts
+}
+
+func (sa *SelfAttention) Forward(ctx ml.Context, hiddenState, positions ml.Tensor, cache kvcache.Cache, opts *Options, isSWA bool) ml.Tensor {
+	batchSize := hiddenState.Dim(1)
+	headDim := cmp.Or(opts.headDim, opts.hiddenSize/opts.numHeads)
+	ropeDim := cmp.Or(opts.ropeDim, headDim)
+
+	query := sa.Query.Forward(ctx, hiddenState)
+	if sa.QNorm != nil {
+		query = sa.QNorm.Forward(ctx, query, opts.eps)
+	}
+	query = query.Reshape(ctx, headDim, opts.numHeads, batchSize)
+
+	key := sa.Key.Forward(ctx, hiddenState)
+	if sa.KNorm != nil {
+		key = sa.KNorm.Forward(ctx, key, opts.eps)
+	}
+	key = key.Reshape(ctx, headDim, opts.numKVHeads, batchSize)
+
+	value := sa.Value.Forward(ctx, hiddenState)
+	value = value.Reshape(ctx, headDim, opts.numKVHeads, batchSize)
+
+	freqScale := float32(1.0)
+	if !isSWA {
+		freqScale = 1. / opts.ropeScale
+	}
+
+	ropeOpts := opts.ropeOptions(sa.RopeFactors, isSWA)
+	query = fast.RoPE(ctx, query, positions, ropeDim, opts.ropeBase, freqScale, ropeOpts...)
+	key = fast.RoPE(ctx, key, positions, ropeDim, opts.ropeBase, freqScale, ropeOpts...)
+
+	attention := nn.Attention(ctx, query, key, value, 1.0/math.Sqrt(float64(headDim)), cache)
+	attention = attention.Reshape(ctx, headDim*opts.numHeads, batchSize)
+	return sa.Output.Forward(ctx, attention)
+}
+
+func (m *Model) Shift(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error) {
+	ropeDim := cmp.Or(m.ropeDim, m.hiddenSize/m.numHeads)
+	isSWA := m.isSWALayer(layer)
+
+	freqScale := float32(1.0)
+	if !isSWA {
+		freqScale = 1. / m.ropeScale
+	}
+
+	ropeOpts := m.Options.ropeOptions(m.Layers[layer].SelfAttention.RopeFactors, isSWA)
+	return fast.RoPE(ctx, key, shift, ropeDim, m.ropeBase, freqScale, ropeOpts...), nil
+}
+
+type MLP struct {
+	Up   *nn.Linear `gguf:"ffn_up"`
+	Down *nn.Linear `gguf:"ffn_down"`
+	Gate *nn.Linear `gguf:"ffn_gate"`
+}
+
+func (mlp *MLP) Forward(ctx ml.Context, hiddenState ml.Tensor, opts *Options) ml.Tensor {
+	hiddenState = mlp.Gate.Forward(ctx, hiddenState).SILU(ctx, mlp.Up.Forward(ctx, hiddenState))
+	return mlp.Down.Forward(ctx, hiddenState)
+}
+
+type Layer struct {
+	SelfAttention     *SelfAttention
+	PostAttentionNorm *nn.RMSNorm `gguf:"post_attention_norm"`
+	MLP               *MLP
+	PostFFWNorm       *nn.RMSNorm `gguf:"post_ffw_norm"`
+}
+
+func (l *Layer) Forward(ctx ml.Context, hiddenState, positions, outputs ml.Tensor, cache kvcache.Cache, opts *Options, isSWA bool) ml.Tensor {
+	residual := hiddenState
+
+	hiddenState = l.SelfAttention.Forward(ctx, hiddenState, positions, cache, opts, isSWA)
+
+	if outputs != nil {
+		hiddenState = hiddenState.Rows(ctx, outputs)
+		residual = residual.Rows(ctx, outputs)
+	}
+
+	if l.PostAttentionNorm != nil {
+		hiddenState = l.PostAttentionNorm.Forward(ctx, hiddenState, opts.eps)
+	}
+
+	ffnInput := hiddenState.Add(ctx, residual)
+
+	hiddenState = l.MLP.Forward(ctx, ffnInput, opts)
+
+	if l.PostFFWNorm != nil {
+		hiddenState = l.PostFFWNorm.Forward(ctx, hiddenState, opts.eps)
+	}
+
+	return hiddenState.Add(ctx, ffnInput)
+}
+
+// isSWALayer returns true if the layer uses sliding window attention.
+// Uses the sliding_window_pattern from the model config if available,
+// otherwise falls back to the default OLMo3 pattern (every 4th layer is full attention).
+func (m *Model) isSWALayer(layerIdx int) bool {
+	if len(m.slidingWindowPattern) > layerIdx {
+		return m.slidingWindowPattern[layerIdx]
+	}
+	// Fallback: OLMo3 pattern where every 4th layer (indices 3, 7, 11, ...) uses full attention
+	return (layerIdx+1)%4 != 0
+}
+
+func (m *Model) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, error) {
+	positions := ctx.Input().FromInts(batch.Positions, len(batch.Positions))
+
+	hiddenState := m.TokenEmbedding.Forward(ctx, batch.Inputs)
+
+	for i, layer := range m.Layers {
+		m.Cache.SetLayer(i)
+
+		isSWA := m.isSWALayer(i)
+
+		// Set cache type for interleaved SWA (OLMo3)
+		if wc, ok := m.Cache.(*kvcache.WrapperCache); ok {
+			if isSWA {
+				wc.SetLayerType(0) // SWA cache
+			} else {
+				wc.SetLayerType(1) // Causal cache
+			}
+		}
+
+		var outputs ml.Tensor
+		if i == len(m.Layers)-1 {
+			outputs = batch.Outputs
+		}
+
+		hiddenState = layer.Forward(ctx, hiddenState, positions, outputs, m.Cache, &m.Options, isSWA)
+	}
+
+	hiddenState = m.OutputNorm.Forward(ctx, hiddenState, m.eps)
+	return m.Output.Forward(ctx, hiddenState), nil
+}
+
+func init() {
+	model.Register("olmo2", New)
+}
--- a/model/models/olmo/testolmo.go
+++ b/model/models/olmo/testolmo.go
@@ -0,0 +1,132 @@
+package olmo
+
+import (
+	"context"
+	"fmt"
+	"log"
+
+	"github.com/ollama/ollama/ml"
+	"github.com/ollama/ollama/model"
+	"github.com/ollama/ollama/model/input"
+	"github.com/ollama/ollama/sample"
+)
+
+func main() {
+	modelPath := "/Users/nicole/models/Olmo-3-7B-Think/olmo-3-7b-think-q8_0.gguf"
+
+	fmt.Println("Loading OLMo model...")
+	m, err := model.New(modelPath, ml.BackendParams{AllocMemory: true})
+	if err != nil {
+		log.Fatal(err)
+	}
+
+	if err := m.Backend().Load(context.Background(), func(f float32) {}); err != nil {
+		log.Fatal(err)
+	}
+
+	fmt.Println("✅ Model loaded successfully!")
+
+	// Initialize the cache
+	cache := m.Config().Cache
+	if cache != nil {
+		// Initialize with reasonable defaults:
+		// - dtype: F16
+		// - maxSequences: 1 (single sequence)
+		// - capacity: 2048 (context length)
+		// - maxBatch: 512
+		cache.Init(m.Backend(), ml.DTypeF16, 1, 2048, 512)
+		fmt.Printf("✅ Cache initialized (type: %T)\n", cache)
+	}
+
+	// Test generation
+	prompt := "Question: What is machine learning? Answer:"
+	fmt.Printf("\nPrompt: %s\n", prompt)
+
+	tp := m.(model.TextProcessor)
+	tokens, err := tp.Encode(prompt, true)
+	if err != nil {
+		log.Fatal(err)
+	}
+
+	fmt.Printf("Tokens: %v (count: %d)\n", tokens, len(tokens))
+
+	// Generate 20 tokens
+	maxTokens := 20
+	generated := make([]int32, 0, maxTokens)
+
+	// Create sampler (temperature=0 for greedy sampling)
+	sampler := sample.NewSampler(0, 0, 0, 0, -1, nil)
+
+	for i := 0; i < maxTokens; i++ {
+		// Create a new context for each generation step to avoid memory buildup
+		ctx := m.Backend().NewContext()
+
+		var inputTokens []int32
+		var positions []int32
+
+		if i == 0 {
+			// First iteration: process all prompt tokens
+			inputTokens = tokens
+			positions = make([]int32, len(tokens))
+			for j := range positions {
+				positions[j] = int32(j)
+			}
+		} else {
+			// Subsequent iterations: only process the newly generated token
+			// The last token is at position len(tokens)-1 (its index in the sequence)
+			inputTokens = []int32{tokens[len(tokens)-1]}
+			positions = []int32{int32(len(tokens) - 1)}
+		}
+
+		sequences := make([]int, len(inputTokens))
+		// All tokens belong to sequence 0
+
+		inputsTensor := ctx.Input().FromInts(inputTokens, len(inputTokens))
+		outputs := ctx.Input().FromInts([]int32{int32(len(inputTokens) - 1)}, 1)
+
+		batch := input.Batch{
+			Inputs:    inputsTensor,
+			Positions: positions,
+			Sequences: sequences,
+			Outputs:   outputs,
+		}
+
+		// Forward pass (model.Forward handles cache.StartForward internally)
+		logits, err := model.Forward(ctx, m, batch)
+		if err != nil {
+			ctx.Close()
+			log.Fatal(err)
+		}
+
+		logits = logits.Contiguous(ctx)
+		ctx.Forward(logits).Compute(logits)
+
+		logitValues := logits.Floats()
+
+		// Sample next token
+		nextToken, err := sampler.Sample(logitValues)
+		if err != nil {
+			ctx.Close()
+			log.Fatal(err)
+		}
+
+		// Close context before moving to next iteration
+		ctx.Close()
+
+		generated = append(generated, nextToken)
+		tokens = append(tokens, nextToken)
+
+		// Decode and print
+		decoded, _ := tp.Decode([]int32{nextToken})
+		fmt.Print(decoded)
+
+		// Stop on EOS
+		if nextToken == 2 || nextToken == 1 { // Common EOS tokens
+			break
+		}
+	}
+
+	fmt.Println("\n\n✅ Generation completed!")
+	fullText, _ := tp.Decode(generated)
+	fmt.Printf("Generated: %s\n", fullText)
+}
--- a/model/parsers/olmo3.go
+++ b/model/parsers/olmo3.go
@@ -0,0 +1,469 @@
+package parsers
+
+import (
+	"context"
+	"fmt"
+	"log/slog"
+	"regexp"
+	"strconv"
+	"strings"
+
+	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/logutil"
+)
+
+type olmo3ParserState int
+
+const (
+	olmo3StateContent olmo3ParserState = iota
+	olmo3StateToolCalls
+	olmo3StateToolCallsDone
+)
+
+const (
+	olmo3FuncCallsOpenTag  = "<function_calls>"
+	olmo3FuncCallsCloseTag = "</function_calls>"
+)
+
+type Olmo3Parser struct {
+	state  olmo3ParserState
+	buffer strings.Builder
+}
+
+func (p *Olmo3Parser) HasToolSupport() bool {
+	return true
+}
+
+func (p *Olmo3Parser) HasThinkingSupport() bool {
+	return false
+}
+
+func (p *Olmo3Parser) Init(tools []api.Tool, lastMessage *api.Message, thinkValue *api.ThinkValue) []api.Tool {
+	p.state = olmo3StateContent
+	return tools
+}
+
+type olmo3ParserEvent interface {
+	isOlmo3ParserEvent()
+}
+
+type olmo3ParserEventContent struct {
+	content string
+}
+
+type olmo3ParserEventToolCalls struct {
+	calls []api.ToolCall
+}
+
+func (olmo3ParserEventContent) isOlmo3ParserEvent()   {}
+func (olmo3ParserEventToolCalls) isOlmo3ParserEvent() {}
+
+func (p *Olmo3Parser) Add(s string, done bool) (content string, thinking string, calls []api.ToolCall, err error) {
+	p.buffer.WriteString(s)
+
+	if done {
+		// Drain any remaining content
+		bufStr := p.buffer.String()
+		p.buffer.Reset()
+		if p.state == olmo3StateContent && len(bufStr) > 0 {
+			return bufStr, "", nil, nil
+		}
+		return "", "", nil, nil
+	}
+
+	events := p.parseEvents()
+
+	var contentSb strings.Builder
+	var allCalls []api.ToolCall
+	for _, event := range events {
+		switch event := event.(type) {
+		case olmo3ParserEventContent:
+			contentSb.WriteString(event.content)
+		case olmo3ParserEventToolCalls:
+			allCalls = append(allCalls, event.calls...)
+		}
+	}
+
+	return contentSb.String(), "", allCalls, nil
+}
+
+func (p *Olmo3Parser) parseEvents() []olmo3ParserEvent {
+	var all []olmo3ParserEvent
+
+	keepLooping := true
+	for keepLooping {
+		var events []olmo3ParserEvent
+		events, keepLooping = p.eat()
+		if len(events) > 0 {
+			all = append(all, events...)
+		}
+	}
+
+	if len(all) > 0 {
+		slog.Log(context.TODO(), logutil.LevelTrace, "olmo3 events parsed", "events", all, "state", p.state, "buffer", p.buffer.String())
+	}
+
+	return all
+}
+
+func (p *Olmo3Parser) eat() ([]olmo3ParserEvent, bool) {
+	var events []olmo3ParserEvent
+	bufStr := p.buffer.String()
+	if bufStr == "" {
+		return events, false
+	}
+
+	switch p.state {
+	case olmo3StateContent:
+		if strings.Contains(bufStr, olmo3FuncCallsOpenTag) {
+			// Found <function_calls> tag
+			split := strings.SplitN(bufStr, olmo3FuncCallsOpenTag, 2)
+			content := split[0]
+			remaining := split[1]
+
+			p.buffer.Reset()
+			p.buffer.WriteString(remaining)
+			p.state = olmo3StateToolCalls
+
+			if len(content) > 0 {
+				events = append(events, olmo3ParserEventContent{content: content})
+			}
+			return events, true
+		} else if overlapLen := overlap(bufStr, olmo3FuncCallsOpenTag); overlapLen > 0 {
+			// Partial <function_calls> tag - withhold ambiguous content
+			unambiguous := bufStr[:len(bufStr)-overlapLen]
+			ambiguous := bufStr[len(bufStr)-overlapLen:]
+			p.buffer.Reset()
+			p.buffer.WriteString(ambiguous)
+			if len(unambiguous) > 0 {
+				events = append(events, olmo3ParserEventContent{content: unambiguous})
+			}
+			return events, false
+		} else {
+			// Regular content - emit all
+			p.buffer.Reset()
+			if len(bufStr) > 0 {
+				events = append(events, olmo3ParserEventContent{content: bufStr})
+			}
+			return events, false
+		}
+
+	case olmo3StateToolCalls:
+		if strings.Contains(bufStr, olmo3FuncCallsCloseTag) {
+			// Found </function_calls> tag
+			split := strings.SplitN(bufStr, olmo3FuncCallsCloseTag, 2)
+			toolCallsStr := split[0]
+			remaining := split[1]
+
+			p.buffer.Reset()
+			p.buffer.WriteString(remaining)
+			p.state = olmo3StateToolCallsDone
+
+			// Parse the function calls
+			calls, err := parseOlmo3FunctionCalls(toolCallsStr)
+			if err != nil {
+				slog.Log(context.TODO(), logutil.LevelTrace, "failed to parse olmo3 function calls", "error", err, "content", toolCallsStr)
+			} else if len(calls) > 0 {
+				events = append(events, olmo3ParserEventToolCalls{calls: calls})
+			}
+			return events, true
+		} else if overlapLen := overlap(bufStr, olmo3FuncCallsCloseTag); overlapLen > 0 {
+			// Partial </function_calls> tag - wait for more
+			return events, false
+		}
+		// Still collecting tool calls, wait for close tag
+		return events, false
+
+	case olmo3StateToolCallsDone:
+		// After tool calls, emit remaining content
+		p.buffer.Reset()
+		p.state = olmo3StateContent
+		if len(bufStr) > 0 {
+			events = append(events, olmo3ParserEventContent{content: bufStr})
+		}
+		return events, false
+	}
+
+	return events, false
+}
+
+// parseOlmo3FunctionCalls parses function calls in Python-esque format:
+// func_name(arg1="value1", arg2=123)
+// Multiple calls are separated by newlines
+func parseOlmo3FunctionCalls(s string) ([]api.ToolCall, error) {
+	var calls []api.ToolCall
+	s = strings.TrimSpace(s)
+	if s == "" {
+		return calls, nil
+	}
+
+	// Split by newlines for multiple function calls
+	lines := strings.Split(s, "\n")
+	for _, line := range lines {
+		line = strings.TrimSpace(line)
+		if line == "" {
+			continue
+		}
+
+		call, err := parseOlmo3SingleFunctionCall(line)
+		if err != nil {
+			return nil, fmt.Errorf("failed to parse function call %q: %w", line, err)
+		}
+		calls = append(calls, call)
+	}
+
+	return calls, nil
+}
+
+// Regex to match function call: func_name(args)
+var funcCallRegex = regexp.MustCompile(`^(\w+)\((.*)\)$`)
+
+// Regex to match a single argument: key=value
+// Value can be: "string", 'string', number, true, false, null, or nested structures
+var argRegex = regexp.MustCompile(`^(\w+)=(.+)$`)
+
+func parseOlmo3SingleFunctionCall(s string) (api.ToolCall, error) {
+	matches := funcCallRegex.FindStringSubmatch(s)
+	if matches == nil {
+		return api.ToolCall{}, fmt.Errorf("invalid function call format")
+	}
+
+	funcName := matches[1]
+	argsStr := matches[2]
+
+	args, err := parseOlmo3Arguments(argsStr)
+	if err != nil {
+		return api.ToolCall{}, fmt.Errorf("failed to parse arguments: %w", err)
+	}
+
+	return api.ToolCall{
+		Function: api.ToolCallFunction{
+			Name:      funcName,
+			Arguments: args,
+		},
+	}, nil
+}
+
+// parseOlmo3Arguments parses comma-separated key=value pairs
+// Handles nested parentheses, brackets, braces, and quoted strings
+func parseOlmo3Arguments(s string) (map[string]any, error) {
+	args := make(map[string]any)
+	s = strings.TrimSpace(s)
+	if s == "" {
+		return args, nil
+	}
+
+	// Split by commas, but respect nested structures and quotes
+	parts := splitArguments(s)
+
+	for _, part := range parts {
+		part = strings.TrimSpace(part)
+		if part == "" {
+			continue
+		}
+
+		// Find the first = sign
+		eqIdx := strings.Index(part, "=")
+		if eqIdx == -1 {
+			return nil, fmt.Errorf("invalid argument format: %s", part)
+		}
+
+		key := strings.TrimSpace(part[:eqIdx])
+		valueStr := strings.TrimSpace(part[eqIdx+1:])
+
+		value, err := parseOlmo3Value(valueStr)
+		if err != nil {
+			return nil, fmt.Errorf("failed to parse value for %s: %w", key, err)
+		}
+
+		args[key] = value
+	}
+
+	return args, nil
+}
+
+// splitArguments splits arguments by commas, respecting quotes and nested structures
+func splitArguments(s string) []string {
+	var parts []string
+	var current strings.Builder
+	depth := 0
+	inString := false
+	stringChar := byte(0)
+	escaped := false
+
+	for i := 0; i < len(s); i++ {
+		c := s[i]
+
+		if escaped {
+			current.WriteByte(c)
+			escaped = false
+			continue
+		}
+
+		if c == '\\' && inString {
+			current.WriteByte(c)
+			escaped = true
+			continue
+		}
+
+		if (c == '"' || c == '\'') && !inString {
+			inString = true
+			stringChar = c
+			current.WriteByte(c)
+			continue
+		}
+
+		if c == stringChar && inString {
+			inString = false
+			stringChar = 0
+			current.WriteByte(c)
+			continue
+		}
+
+		if !inString {
+			switch c {
+			case '(', '[', '{':
+				depth++
+				current.WriteByte(c)
+			case ')', ']', '}':
+				depth--
+				current.WriteByte(c)
+			case ',':
+				if depth == 0 {
+					parts = append(parts, current.String())
+					current.Reset()
+					continue
+				}
+				current.WriteByte(c)
+			default:
+				current.WriteByte(c)
+			}
+		} else {
+			current.WriteByte(c)
+		}
+	}
+
+	if current.Len() > 0 {
+		parts = append(parts, current.String())
+	}
+
+	return parts
+}
+
+// parseOlmo3Value parses a value which can be a string, number, boolean, null, array, or object
+func parseOlmo3Value(s string) (any, error) {
+	s = strings.TrimSpace(s)
+
+	// Check for quoted string
+	if (strings.HasPrefix(s, `"`) && strings.HasSuffix(s, `"`)) ||
+		(strings.HasPrefix(s, `'`) && strings.HasSuffix(s, `'`)) {
+		// Remove quotes and unescape
+		inner := s[1 : len(s)-1]
+		return unescapeString(inner), nil
+	}
+
+	// Check for boolean
+	if s == "true" || s == "True" {
+		return true, nil
+	}
+	if s == "false" || s == "False" {
+		return false, nil
+	}
+
+	// Check for null/None
+	if s == "null" || s == "None" || s == "nil" {
+		return nil, nil
+	}
+
+	// Check for number
+	if i, err := strconv.ParseInt(s, 10, 64); err == nil {
+		return i, nil
+	}
+	if f, err := strconv.ParseFloat(s, 64); err == nil {
+		return f, nil
+	}
+
+	// Check for array [...]
+	if strings.HasPrefix(s, "[") && strings.HasSuffix(s, "]") {
+		return parseOlmo3Array(s[1 : len(s)-1])
+	}
+
+	// Check for object {...}
+	if strings.HasPrefix(s, "{") && strings.HasSuffix(s, "}") {
+		return parseOlmo3Object(s[1 : len(s)-1])
+	}
+
+	// Default to string without quotes
+	return s, nil
+}
+
+func parseOlmo3Array(s string) ([]any, error) {
+	s = strings.TrimSpace(s)
+	if s == "" {
+		return []any{}, nil
+	}
+
+	parts := splitArguments(s)
+	var arr []any
+	for _, part := range parts {
+		val, err := parseOlmo3Value(part)
+		if err != nil {
+			return nil, err
+		}
+		arr = append(arr, val)
+	}
+	return arr, nil
+}
+
+func parseOlmo3Object(s string) (map[string]any, error) {
+	s = strings.TrimSpace(s)
+	if s == "" {
+		return map[string]any{}, nil
+	}
+
+	// Objects use key: value or "key": value format
+	obj := make(map[string]any)
+	parts := splitArguments(s)
+	for _, part := range parts {
+		part = strings.TrimSpace(part)
+		if part == "" {
+			continue
+		}
+
+		// Find colon separator
+		colonIdx := strings.Index(part, ":")
+		if colonIdx == -1 {
+			return nil, fmt.Errorf("invalid object entry: %s", part)
+		}
+
+		keyStr := strings.TrimSpace(part[:colonIdx])
+		valueStr := strings.TrimSpace(part[colonIdx+1:])
+
+		// Remove quotes from key if present
+		if (strings.HasPrefix(keyStr, `"`) && strings.HasSuffix(keyStr, `"`)) ||
+			(strings.HasPrefix(keyStr, `'`) && strings.HasSuffix(keyStr, `'`)) {
+			keyStr = keyStr[1 : len(keyStr)-1]
+		}
+
+		val, err := parseOlmo3Value(valueStr)
+		if err != nil {
+			return nil, fmt.Errorf("failed to parse value for key %s: %w", keyStr, err)
+		}
+
+		obj[keyStr] = val
+	}
+
+	return obj, nil
+}
+
+func unescapeString(s string) string {
+	// Handle common escape sequences
+	s = strings.ReplaceAll(s, `\\`, "\x00") // Placeholder for backslash
+	s = strings.ReplaceAll(s, `\"`, `"`)
+	s = strings.ReplaceAll(s, `\'`, `'`)
+	s = strings.ReplaceAll(s, `\n`, "\n")
+	s = strings.ReplaceAll(s, `\t`, "\t")
+	s = strings.ReplaceAll(s, `\r`, "\r")
+	s = strings.ReplaceAll(s, "\x00", `\`) // Restore backslash
+	return s
+}
--- a/model/parsers/olmo3_test.go
+++ b/model/parsers/olmo3_test.go
@@ -0,0 +1,483 @@
+package parsers
+
+import (
+	"testing"
+
+	"github.com/google/go-cmp/cmp"
+
+	"github.com/ollama/ollama/api"
+)
+
+func TestOlmo3Parser(t *testing.T) {
+	tests := []struct {
+		name             string
+		input            string
+		expectedContent  string
+		expectedThinking string
+		expectedCalls    []api.ToolCall
+	}{
+		{
+			name:            "simple content",
+			input:           "Hello, how can I help you?",
+			expectedContent: "Hello, how can I help you?",
+		},
+		{
+			name:  "simple tool call",
+			input: `<function_calls>get_weather(location="San Francisco")</function_calls>`,
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "get_weather",
+						Arguments: map[string]any{"location": "San Francisco"},
+					},
+				},
+			},
+		},
+		{
+			name:            "content then tool call",
+			input:           `Let me check the weather.<function_calls>get_weather(location="NYC")</function_calls>`,
+			expectedContent: "Let me check the weather.",
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "get_weather",
+						Arguments: map[string]any{"location": "NYC"},
+					},
+				},
+			},
+		},
+		{
+			name:  "tool call with multiple arguments",
+			input: `<function_calls>book_flight(from="SFO", to="NYC", date="2024-01-15")</function_calls>`,
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name: "book_flight",
+						Arguments: map[string]any{
+							"from": "SFO",
+							"to":   "NYC",
+							"date": "2024-01-15",
+						},
+					},
+				},
+			},
+		},
+		{
+			name: "multiple tool calls",
+			input: `<function_calls>get_weather(location="San Francisco")
+get_weather(location="New York")</function_calls>`,
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "get_weather",
+						Arguments: map[string]any{"location": "San Francisco"},
+					},
+				},
+				{
+					Function: api.ToolCallFunction{
+						Name:      "get_weather",
+						Arguments: map[string]any{"location": "New York"},
+					},
+				},
+			},
+		},
+		{
+			name:  "tool call with numeric argument",
+			input: `<function_calls>set_temperature(value=72)</function_calls>`,
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "set_temperature",
+						Arguments: map[string]any{"value": int64(72)},
+					},
+				},
+			},
+		},
+		{
+			name:  "tool call with float argument",
+			input: `<function_calls>set_price(amount=19.99)</function_calls>`,
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "set_price",
+						Arguments: map[string]any{"amount": 19.99},
+					},
+				},
+			},
+		},
+		{
+			name:  "tool call with boolean argument",
+			input: `<function_calls>toggle_setting(enabled=true)</function_calls>`,
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "toggle_setting",
+						Arguments: map[string]any{"enabled": true},
+					},
+				},
+			},
+		},
+		{
+			name:  "tool call with null argument",
+			input: `<function_calls>clear_value(field=null)</function_calls>`,
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "clear_value",
+						Arguments: map[string]any{"field": nil},
+					},
+				},
+			},
+		},
+		{
+			name:  "tool call with array argument",
+			input: `<function_calls>process_items(items=["apple", "banana", "cherry"])</function_calls>`,
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "process_items",
+						Arguments: map[string]any{"items": []any{"apple", "banana", "cherry"}},
+					},
+				},
+			},
+		},
+		{
+			name:  "tool call with dict argument",
+			input: `<function_calls>update_config(settings={"theme": "dark", "fontSize": 14})</function_calls>`,
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name: "update_config",
+						Arguments: map[string]any{
+							"settings": map[string]any{
+								"theme":    "dark",
+								"fontSize": int64(14),
+							},
+						},
+					},
+				},
+			},
+		},
+		{
+			name:  "tool call with nested dict",
+			input: `<function_calls>create_request(data={"user": {"name": "John", "age": 30}, "active": true})</function_calls>`,
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name: "create_request",
+						Arguments: map[string]any{
+							"data": map[string]any{
+								"user": map[string]any{
+									"name": "John",
+									"age":  int64(30),
+								},
+								"active": true,
+							},
+						},
+					},
+				},
+			},
+		},
+		{
+			name:  "tool call with no arguments",
+			input: `<function_calls>get_current_time()</function_calls>`,
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "get_current_time",
+						Arguments: map[string]any{},
+					},
+				},
+			},
+		},
+		{
+			name:  "tool call with single quotes",
+			input: `<function_calls>search(query='hello world')</function_calls>`,
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "search",
+						Arguments: map[string]any{"query": "hello world"},
+					},
+				},
+			},
+		},
+		{
+			name:  "tool call with escaped quotes",
+			input: `<function_calls>search(query="say \"hello\"")</function_calls>`,
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "search",
+						Arguments: map[string]any{"query": `say "hello"`},
+					},
+				},
+			},
+		},
+		{
+			name:  "tool call with mixed argument types",
+			input: `<function_calls>create_user(name="John", age=30, active=true)</function_calls>`,
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name: "create_user",
+						Arguments: map[string]any{
+							"name":   "John",
+							"age":    int64(30),
+							"active": true,
+						},
+					},
+				},
+			},
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			p := &Olmo3Parser{}
+			p.Init(nil, nil, nil)
+
+			content, thinking, calls, err := p.Add(tt.input, false)
+			if err != nil {
+				t.Fatalf("unexpected error: %v", err)
+			}
+
+			// Drain remaining content
+			finalContent, finalThinking, finalCalls, err := p.Add("", true)
+			if err != nil {
+				t.Fatalf("unexpected error on done: %v", err)
+			}
+			content += finalContent
+			thinking += finalThinking
+			calls = append(calls, finalCalls...)
+
+			if diff := cmp.Diff(content, tt.expectedContent); diff != "" {
+				t.Errorf("content mismatch (-got +want):\n%s", diff)
+			}
+			if diff := cmp.Diff(thinking, tt.expectedThinking); diff != "" {
+				t.Errorf("thinking mismatch (-got +want):\n%s", diff)
+			}
+			if diff := cmp.Diff(calls, tt.expectedCalls); diff != "" {
+				t.Errorf("calls mismatch (-got +want):\n%s", diff)
+			}
+		})
+	}
+}
+
+func TestOlmo3Parser_Streaming(t *testing.T) {
+	tests := []struct {
+		name            string
+		chunks          []string
+		expectedContent string
+		expectedCalls   []api.ToolCall
+	}{
+		{
+			name:            "streaming content",
+			chunks:          []string{"Hello, ", "how ", "can I help?"},
+			expectedContent: "Hello, how can I help?",
+		},
+		{
+			name:   "streaming tool call",
+			chunks: []string{"<function_", "calls>get_weather", "(location=\"SF\")", "</function_calls>"},
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "get_weather",
+						Arguments: map[string]any{"location": "SF"},
+					},
+				},
+			},
+		},
+		{
+			name:            "streaming content then tool call",
+			chunks:          []string{"Let me check.", "<function_calls>", "get_weather(location=\"NYC\")", "</function_calls>"},
+			expectedContent: "Let me check.",
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "get_weather",
+						Arguments: map[string]any{"location": "NYC"},
+					},
+				},
+			},
+		},
+		{
+			name:   "tool call tag split across chunks",
+			chunks: []string{"<func", "tion_calls>test()</function_calls>"},
+			expectedCalls: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "test",
+						Arguments: map[string]any{},
+					},
+				},
+			},
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			p := &Olmo3Parser{}
+			p.Init(nil, nil, nil)
+
+			var allContent string
+			var allCalls []api.ToolCall
+
+			for _, chunk := range tt.chunks {
+				content, _, calls, err := p.Add(chunk, false)
+				if err != nil {
+					t.Fatalf("unexpected error: %v", err)
+				}
+				allContent += content
+				allCalls = append(allCalls, calls...)
+			}
+
+			// Drain
+			content, _, calls, err := p.Add("", true)
+			if err != nil {
+				t.Fatalf("unexpected error on done: %v", err)
+			}
+			allContent += content
+			allCalls = append(allCalls, calls...)
+
+			if diff := cmp.Diff(allContent, tt.expectedContent); diff != "" {
+				t.Errorf("content mismatch (-got +want):\n%s", diff)
+			}
+			if diff := cmp.Diff(allCalls, tt.expectedCalls); diff != "" {
+				t.Errorf("calls mismatch (-got +want):\n%s", diff)
+			}
+		})
+	}
+}
+
+func TestOlmo3Parser_HasToolSupport(t *testing.T) {
+	p := &Olmo3Parser{}
+	if !p.HasToolSupport() {
+		t.Error("expected HasToolSupport to return true")
+	}
+}
+
+func TestOlmo3Parser_HasThinkingSupport(t *testing.T) {
+	p := &Olmo3Parser{}
+	if p.HasThinkingSupport() {
+		t.Error("expected HasThinkingSupport to return false")
+	}
+}
+
+func TestParseOlmo3FunctionCalls(t *testing.T) {
+	tests := []struct {
+		name     string
+		input    string
+		expected []api.ToolCall
+		wantErr  bool
+	}{
+		{
+			name:  "simple call",
+			input: `get_weather(location="SF")`,
+			expected: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "get_weather",
+						Arguments: map[string]any{"location": "SF"},
+					},
+				},
+			},
+		},
+		{
+			name:  "multiple args",
+			input: `send_email(to="user@example.com", subject="Hello", body="Test message")`,
+			expected: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name: "send_email",
+						Arguments: map[string]any{
+							"to":      "user@example.com",
+							"subject": "Hello",
+							"body":    "Test message",
+						},
+					},
+				},
+			},
+		},
+		{
+			name: "multiple calls with newlines",
+			input: `get_weather(location="SF")
+get_time(timezone="PST")`,
+			expected: []api.ToolCall{
+				{
+					Function: api.ToolCallFunction{
+						Name:      "get_weather",
+						Arguments: map[string]any{"location": "SF"},
+					},
+				},
+				{
+					Function: api.ToolCallFunction{
+						Name:      "get_time",
+						Arguments: map[string]any{"timezone": "PST"},
+					},
+				},
+			},
+		},
+		{
+			name:     "empty input",
+			input:    "",
+			expected: nil,
+		},
+		{
+			name:     "whitespace only",
+			input:    "   \n   ",
+			expected: nil,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			calls, err := parseOlmo3FunctionCalls(tt.input)
+			if (err != nil) != tt.wantErr {
+				t.Errorf("parseOlmo3FunctionCalls() error = %v, wantErr %v", err, tt.wantErr)
+				return
+			}
+			if diff := cmp.Diff(calls, tt.expected); diff != "" {
+				t.Errorf("calls mismatch (-got +want):\n%s", diff)
+			}
+		})
+	}
+}
+
+func TestParseOlmo3Value(t *testing.T) {
+	tests := []struct {
+		name     string
+		input    string
+		expected any
+	}{
+		{"string double quotes", `"hello"`, "hello"},
+		{"string single quotes", `'hello'`, "hello"},
+		{"integer", "42", int64(42)},
+		{"negative integer", "-10", int64(-10)},
+		{"float", "3.14", 3.14},
+		{"boolean true", "true", true},
+		{"boolean True", "True", true},
+		{"boolean false", "false", false},
+		{"null", "null", nil},
+		{"None", "None", nil},
+		{"empty array", "[]", []any{}},
+		{"array with strings", `["a", "b"]`, []any{"a", "b"}},
+		{"array with numbers", "[1, 2, 3]", []any{int64(1), int64(2), int64(3)}},
+		{"empty object", "{}", map[string]any{}},
+		{"simple object", `{"name": "John"}`, map[string]any{"name": "John"}},
+		{"object with number", `{"age": 30}`, map[string]any{"age": int64(30)}},
+		{"object with multiple keys", `{"a": 1, "b": 2}`, map[string]any{"a": int64(1), "b": int64(2)}},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			result, err := parseOlmo3Value(tt.input)
+			if err != nil {
+				t.Fatalf("unexpected error: %v", err)
+			}
+			if diff := cmp.Diff(result, tt.expected); diff != "" {
+				t.Errorf("value mismatch (-got +want):\n%s", diff)
+			}
+		})
+	}
+}
--- a/model/parsers/parsers.go
+++ b/model/parsers/parsers.go
@@ -54,6 +54,8 @@ func ParserForName(name string) Parser {
 		return harmony.NewHarmonyMessageHandler()
 	case "cogito":
 		return &CogitoParser{}
+	case "olmo3":
+		return &Olmo3Parser{}
 	default:
 		return nil
 	}
--- a/model/renderers/olmo3.go
+++ b/model/renderers/olmo3.go
@@ -0,0 +1,147 @@
+package renderers
+
+import (
+	"encoding/json"
+	"fmt"
+	"sort"
+	"strings"
+
+	"github.com/ollama/ollama/api"
+)
+
+const (
+	olmo3DefaultSystemMessage = "You are a helpful function-calling AI assistant. "
+	olmo3NoFunctionsMessage   = "You do not currently have access to any functions. "
+	olmo3WithFunctionsMessage = "You are provided with function signatures within <functions></functions> XML tags. You may call one or more functions to assist with the user query. Output any function calls within <function_calls></function_calls> XML tags. Do not make assumptions about what values to plug into functions."
+)
+
+type Olmo3Renderer struct{}
+
+func (r *Olmo3Renderer) Render(messages []api.Message, tools []api.Tool, _ *api.ThinkValue) (string, error) {
+	var sb strings.Builder
+
+	var systemMessage *api.Message
+	filteredMessages := make([]api.Message, 0, len(messages))
+	for i, message := range messages {
+		if message.Role == "system" {
+			if systemMessage == nil {
+				systemMessage = &messages[i]
+			}
+			continue
+		}
+		filteredMessages = append(filteredMessages, message)
+	}
+
+	// Render system message
+	if systemMessage != nil {
+		// Custom system message - single newline after "system"
+		sb.WriteString("<|im_start|>system\n")
+		sb.WriteString(systemMessage.Content)
+
+		if len(tools) > 0 {
+			functionsJSON, err := marshalWithSpaces(tools)
+			if err != nil {
+				return "", err
+			}
+			sb.WriteString("<functions>")
+			sb.WriteString(string(functionsJSON))
+			sb.WriteString("</functions>")
+		}
+		sb.WriteString("<|im_end|>\n")
+	} else {
+		// Default system message - single newline after "system"
+		sb.WriteString("<|im_start|>system\n")
+		sb.WriteString(olmo3DefaultSystemMessage)
+
+		if len(tools) > 0 {
+			functionsJSON, err := marshalWithSpaces(tools)
+			if err != nil {
+				return "", err
+			}
+			sb.WriteString(olmo3WithFunctionsMessage)
+			sb.WriteString("<functions>")
+			sb.WriteString(string(functionsJSON))
+			sb.WriteString("</functions>")
+		} else {
+			sb.WriteString(olmo3NoFunctionsMessage)
+			sb.WriteString("<functions></functions>")
+		}
+		sb.WriteString("<|im_end|>\n")
+	}
+
+	for i, message := range filteredMessages {
+		lastMessage := i == len(filteredMessages)-1
+
+		switch message.Role {
+		case "user":
+			sb.WriteString("<|im_start|>user\n")
+			sb.WriteString(message.Content)
+			sb.WriteString("<|im_end|>\n")
+
+		case "assistant":
+			sb.WriteString("<|im_start|>assistant\n")
+
+			if message.Content != "" {
+				sb.WriteString(message.Content)
+			}
+
+			if len(message.ToolCalls) > 0 {
+				sb.WriteString("<function_calls>")
+				for j, tc := range message.ToolCalls {
+					// Format as function_name(arg1="value1", arg2="value2")
+					sb.WriteString(tc.Function.Name)
+					sb.WriteString("(")
+
+					// Get sorted keys for deterministic output
+					keys := make([]string, 0, len(tc.Function.Arguments))
+					for k := range tc.Function.Arguments {
+						keys = append(keys, k)
+					}
+					sort.Strings(keys)
+
+					for k, key := range keys {
+						if k > 0 {
+							sb.WriteString(", ")
+						}
+						value, err := json.Marshal(tc.Function.Arguments[key])
+						if err != nil {
+							return "", err
+						}
+						sb.WriteString(fmt.Sprintf("%s=%s", key, string(value)))
+					}
+					sb.WriteString(")")
+
+					if j < len(message.ToolCalls)-1 {
+						sb.WriteString("\n")
+					}
+				}
+				sb.WriteString("</function_calls>")
+			}
+
+			// Add end tag unless it's the last message with content only (prefill)
+			if !lastMessage || len(message.ToolCalls) > 0 {
+				sb.WriteString("<|im_end|>\n")
+			}
+
+		case "tool":
+			sb.WriteString("<|im_start|>environment\n")
+			sb.WriteString(message.Content)
+			sb.WriteString("<|im_end|>\n")
+		}
+	}
+
+	// Add generation prompt if needed
+	needsGenerationPrompt := true
+	if len(filteredMessages) > 0 {
+		lastMsg := filteredMessages[len(filteredMessages)-1]
+		if lastMsg.Role == "assistant" && len(lastMsg.ToolCalls) == 0 && lastMsg.Content != "" {
+			needsGenerationPrompt = false
+		}
+	}
+
+	if needsGenerationPrompt {
+		sb.WriteString("<|im_start|>assistant\n")
+	}
+
+	return sb.String(), nil
+}
--- a/model/renderers/olmo3_test.go
+++ b/model/renderers/olmo3_test.go
@@ -0,0 +1,290 @@
+package renderers
+
+import (
+	"testing"
+
+	"github.com/google/go-cmp/cmp"
+
+	"github.com/ollama/ollama/api"
+)
+
+func TestOlmo3Renderer(t *testing.T) {
+	tests := []struct {
+		name     string
+		msgs     []api.Message
+		tools    []api.Tool
+		expected string
+	}{
+		{
+			name: "basic without system - adds default system",
+			msgs: []api.Message{
+				{Role: "user", Content: "Hello!"},
+			},
+			expected: "<|im_start|>system\n" +
+				"You are a helpful function-calling AI assistant. You do not currently have access to any functions. <functions></functions><|im_end|>\n" +
+				"<|im_start|>user\n" +
+				"Hello!<|im_end|>\n" +
+				"<|im_start|>assistant\n\n",
+		},
+		{
+			name: "with system message no tools",
+			msgs: []api.Message{
+				{Role: "system", Content: "You are a helpful assistant."},
+				{Role: "user", Content: "Hello!"},
+			},
+			expected: "<|im_start|>system\n" +
+				"You are a helpful assistant.<|im_end|>\n" +
+				"<|im_start|>user\n" +
+				"Hello!<|im_end|>\n" +
+				"<|im_start|>assistant\n\n",
+		},
+		{
+			name: "with system message and tools",
+			msgs: []api.Message{
+				{Role: "system", Content: "You are a helpful assistant."},
+				{Role: "user", Content: "What is the weather?"},
+			},
+			tools: []api.Tool{
+				{
+					Type: "function",
+					Function: api.ToolFunction{
+						Name:        "get_weather",
+						Description: "Get the current weather",
+						Parameters: api.ToolFunctionParameters{
+							Type:     "object",
+							Required: []string{"location"},
+							Properties: map[string]api.ToolProperty{
+								"location": {Type: api.PropertyType{"string"}, Description: "The city"},
+							},
+						},
+					},
+				},
+			},
+			expected: "<|im_start|>system\n" +
+				`You are a helpful assistant.<functions>[{"type": "function", "function": {"name": "get_weather", "description": "Get the current weather", "parameters": {"type": "object", "required": ["location"], "properties": {"location": {"type": "string", "description": "The city"}}}}}]</functions><|im_end|>` + "\n" +
+				"<|im_start|>user\n" +
+				"What is the weather?<|im_end|>\n" +
+				"<|im_start|>assistant\n\n",
+		},
+		{
+			name: "default system with tools - includes function instruction",
+			msgs: []api.Message{
+				{Role: "user", Content: "What is the weather?"},
+			},
+			tools: []api.Tool{
+				{
+					Type: "function",
+					Function: api.ToolFunction{
+						Name:        "get_weather",
+						Description: "Get the current weather",
+						Parameters: api.ToolFunctionParameters{
+							Type:     "object",
+							Required: []string{"location"},
+							Properties: map[string]api.ToolProperty{
+								"location": {Type: api.PropertyType{"string"}, Description: "The city"},
+							},
+						},
+					},
+				},
+			},
+			expected: "<|im_start|>system\n" +
+				"You are a helpful function-calling AI assistant. " +
+				"You are provided with function signatures within <functions></functions> XML tags. You may call one or more functions to assist with the user query. Output any function calls within <function_calls></function_calls> XML tags. Do not make assumptions about what values to plug into functions." +
+				`<functions>[{"type": "function", "function": {"name": "get_weather", "description": "Get the current weather", "parameters": {"type": "object", "required": ["location"], "properties": {"location": {"type": "string", "description": "The city"}}}}}]</functions><|im_end|>` + "\n" +
+				"<|im_start|>user\n" +
+				"What is the weather?<|im_end|>\n" +
+				"<|im_start|>assistant\n\n",
+		},
+		{
+			name: "assistant with tool calls - function call syntax",
+			msgs: []api.Message{
+				{Role: "system", Content: "You are a helpful assistant."},
+				{Role: "user", Content: "What is the weather in SF?"},
+				{
+					Role:    "assistant",
+					Content: "Let me check the weather.",
+					ToolCalls: []api.ToolCall{
+						{
+							ID: "call_1",
+							Function: api.ToolCallFunction{
+								Name: "get_weather",
+								Arguments: map[string]any{
+									"location": "San Francisco",
+								},
+							},
+						},
+					},
+				},
+				{Role: "tool", Content: `{"temperature": 68}`, ToolName: "get_weather"},
+			},
+			tools: []api.Tool{
+				{
+					Type: "function",
+					Function: api.ToolFunction{
+						Name:        "get_weather",
+						Description: "Get the current weather",
+						Parameters: api.ToolFunctionParameters{
+							Type:     "object",
+							Required: []string{"location"},
+							Properties: map[string]api.ToolProperty{
+								"location": {Type: api.PropertyType{"string"}, Description: "The city"},
+							},
+						},
+					},
+				},
+			},
+			expected: "<|im_start|>system\n" +
+				`You are a helpful assistant.<functions>[{"type": "function", "function": {"name": "get_weather", "description": "Get the current weather", "parameters": {"type": "object", "required": ["location"], "properties": {"location": {"type": "string", "description": "The city"}}}}}]</functions><|im_end|>` + "\n" +
+				"<|im_start|>user\n" +
+				"What is the weather in SF?<|im_end|>\n" +
+				"<|im_start|>assistant\n" +
+				`Let me check the weather.<function_calls>get_weather(location="San Francisco")</function_calls><|im_end|>` + "\n" +
+				"<|im_start|>environment\n" +
+				`{"temperature": 68}<|im_end|>` + "\n" +
+				"<|im_start|>assistant\n\n",
+		},
+		{
+			name: "multi-turn conversation",
+			msgs: []api.Message{
+				{Role: "system", Content: "You are a helpful assistant."},
+				{Role: "user", Content: "Hello"},
+				{Role: "assistant", Content: "Hi there!"},
+				{Role: "user", Content: "How are you?"},
+			},
+			expected: "<|im_start|>system\n" +
+				"You are a helpful assistant.<|im_end|>\n" +
+				"<|im_start|>user\n" +
+				"Hello<|im_end|>\n" +
+				"<|im_start|>assistant\n" +
+				"Hi there!<|im_end|>\n" +
+				"<|im_start|>user\n" +
+				"How are you?<|im_end|>\n" +
+				"<|im_start|>assistant\n\n",
+		},
+		{
+			name: "parallel tool calls - newline separated",
+			msgs: []api.Message{
+				{Role: "user", Content: "Get weather in SF and NYC"},
+				{
+					Role: "assistant",
+					ToolCalls: []api.ToolCall{
+						{
+							ID: "call_1",
+							Function: api.ToolCallFunction{
+								Name:      "get_weather",
+								Arguments: map[string]any{"location": "San Francisco"},
+							},
+						},
+						{
+							ID: "call_2",
+							Function: api.ToolCallFunction{
+								Name:      "get_weather",
+								Arguments: map[string]any{"location": "New York"},
+							},
+						},
+					},
+				},
+				{Role: "tool", Content: `{"temperature": 68}`, ToolName: "get_weather"},
+				{Role: "tool", Content: `{"temperature": 55}`, ToolName: "get_weather"},
+			},
+			tools: []api.Tool{
+				{
+					Type: "function",
+					Function: api.ToolFunction{
+						Name: "get_weather",
+						Parameters: api.ToolFunctionParameters{
+							Type: "object",
+							Properties: map[string]api.ToolProperty{
+								"location": {Type: api.PropertyType{"string"}},
+							},
+						},
+					},
+				},
+			},
+			expected: "<|im_start|>system\n" +
+				"You are a helpful function-calling AI assistant. " +
+				"You are provided with function signatures within <functions></functions> XML tags. You may call one or more functions to assist with the user query. Output any function calls within <function_calls></function_calls> XML tags. Do not make assumptions about what values to plug into functions." +
+				`<functions>[{"type": "function", "function": {"name": "get_weather", "parameters": {"type": "object", "properties": {"location": {"type": "string"}}}}}]</functions><|im_end|>` + "\n" +
+				"<|im_start|>user\n" +
+				"Get weather in SF and NYC<|im_end|>\n" +
+				"<|im_start|>assistant\n" +
+				`<function_calls>get_weather(location="San Francisco")` + "\n" +
+				`get_weather(location="New York")</function_calls><|im_end|>` + "\n" +
+				"<|im_start|>environment\n" +
+				`{"temperature": 68}<|im_end|>` + "\n" +
+				"<|im_start|>environment\n" +
+				`{"temperature": 55}<|im_end|>` + "\n" +
+				"<|im_start|>assistant\n\n",
+		},
+		{
+			name: "tool call with multiple arguments",
+			msgs: []api.Message{
+				{Role: "user", Content: "Book a flight"},
+				{
+					Role: "assistant",
+					ToolCalls: []api.ToolCall{
+						{
+							ID: "call_1",
+							Function: api.ToolCallFunction{
+								Name: "book_flight",
+								Arguments: map[string]any{
+									"from": "SFO",
+									"to":   "NYC",
+								},
+							},
+						},
+					},
+				},
+			},
+			tools: []api.Tool{
+				{
+					Type: "function",
+					Function: api.ToolFunction{
+						Name: "book_flight",
+						Parameters: api.ToolFunctionParameters{
+							Type: "object",
+							Properties: map[string]api.ToolProperty{
+								"from": {Type: api.PropertyType{"string"}},
+								"to":   {Type: api.PropertyType{"string"}},
+							},
+						},
+					},
+				},
+			},
+			expected: "<|im_start|>system\n" +
+				"You are a helpful function-calling AI assistant. " +
+				"You are provided with function signatures within <functions></functions> XML tags. You may call one or more functions to assist with the user query. Output any function calls within <function_calls></function_calls> XML tags. Do not make assumptions about what values to plug into functions." +
+				`<functions>[{"type": "function", "function": {"name": "book_flight", "parameters": {"type": "object", "properties": {"from": {"type": "string"}, "to": {"type": "string"}}}}}]</functions><|im_end|>` + "\n" +
+				"<|im_start|>user\n" +
+				"Book a flight<|im_end|>\n" +
+				"<|im_start|>assistant\n" +
+				`<function_calls>book_flight(from="SFO", to="NYC")</function_calls><|im_end|>` + "\n" +
+				"<|im_start|>assistant\n\n",
+		},
+		{
+			name: "assistant prefill - no generation prompt",
+			msgs: []api.Message{
+				{Role: "user", Content: "Hello"},
+				{Role: "assistant", Content: "Hi there!"},
+			},
+			expected: "<|im_start|>system\n" +
+				"You are a helpful function-calling AI assistant. You do not currently have access to any functions. <functions></functions><|im_end|>\n" +
+				"<|im_start|>user\n" +
+				"Hello<|im_end|>\n" +
+				"<|im_start|>assistant\n" +
+				"Hi there!",
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			rendered, err := (&Olmo3Renderer{}).Render(tt.msgs, tt.tools, nil)
+			if err != nil {
+				t.Fatal(err)
+			}
+			if diff := cmp.Diff(rendered, tt.expected); diff != "" {
+				t.Errorf("mismatch (-got +want):\n%s", diff)
+			}
+		})
+	}
+}
--- a/model/renderers/renderer.go
+++ b/model/renderers/renderer.go
@@ -59,6 +59,9 @@ func rendererForName(name string) Renderer {
 	case "cogito":
 		renderer := &CogitoRenderer{isThinking: true}
 		return renderer
+	case "olmo3":
+		renderer := &Olmo3Renderer{}
+		return renderer
 	default:
 		return nil
 	}
--- a/server/prompt.go
+++ b/server/prompt.go
@@ -110,6 +110,7 @@ func renderPrompt(m *Model, msgs []api.Message, tools []api.Tool, think *api.Thi
 		if err != nil {
 			return "", err
 		}
+		slog.Debug("rendered prompt", "renderer", m.Config.Renderer, "prompt", rendered)
 		return rendered, nil
 	}
Author	SHA1	Message	Date
ParthSareen	5c3bf414ef	close to working	2025-12-08 18:17:56 -08:00
ParthSareen	0a9862a383	wip	2025-12-08 15:20:52 -08:00
ParthSareen	f475cc365a	wip	2025-12-08 14:42:18 -08:00
ParthSareen	dd3306d3a0	renderers/parsers: olmo3 instruct	2025-12-08 13:23:11 -08:00
nicole pardal	57c1d7db9a	fixed generation issue	2025-12-08 00:35:49 -08:00
nicole pardal	91d6370a62	removed original olmo support	2025-12-01 14:17:46 -08:00
nicole pardal	38a2a6468f	removed olmo1 support	2025-12-01 14:14:31 -08:00
nicole pardal	064ec63ddf	lint	2025-11-26 20:05:25 -08:00
nicole pardal	fd959fbf7a	updated converter	2025-11-26 19:42:34 -08:00
nicole pardal	cfc9729edf	olmo model initial	2025-11-25 15:49:09 -08:00