chore(gallery): add Kayak Qwen3-VL-4B variants

Add Q4_K_M and Q8_0 builds with the matching vision projector for
single-step screenshot grounding on Kayak interfaces.

Assisted-by: Codex:gpt-6-astra

docs/content/features/model-gallery.md

@@ -39,6 +39,28 @@ Both views use the same model selection and store the view, search, filter, and
 selection in the URL. Installing from Explore does not move you away from the
 catalog; the entry updates in place when the operation finishes.
 
+## RuntimeLabs Kayak screenshot grounding
+
+[RuntimeLabs Kayak](https://huggingface.co/ayushadarsh7/RuntimeLabs-Kayak-Qwen3-VL-4B)
+is a Qwen3-VL-4B fine-tune for locating controls in Kayak flight-search screenshots.
+The gallery offers Q4_K_M and Q8_0 GGUF variants for llama.cpp, each with the
+matching F16 vision projector and an 8,192-token context.
+
+Install the Q4_K_M build explicitly:
+
+```bash
+local-ai models install runtimelabs-kayak-qwen3-vl-4b-q4 --variant runtimelabs-kayak-qwen3-vl-4b-q4
+```
+
+Send a screenshot and one instruction through `/v1/chat/completions`. Use the
+prompt format in the publisher's [inference script](https://huggingface.co/ayushadarsh7/RuntimeLabs-Kayak-Qwen3-VL-4B/blob/main/infer.py),
+with `temperature: 0` and `max_tokens: 128`. Fill instructions can include the
+text to enter. The response contains an action and, when applicable, a bounding
+box in `[x, y, width, height]` format on a 0–1000 scale; x and y identify its
+top-left corner. The model predicts coordinates; your application executes the
+action. Its training covers Kayak interfaces, so performance on other sites or
+changed layouts is unverified.
+
 ## Cyber-Ornith 1.5 9B
 
 Cyber-Ornith 1.5 is a Qwen3.5 fine-tune for security auditing, terminal tasks, and tool use.

gallery/index.yaml

@@ -22646,6 +22646,94 @@
     - filename: mmproj/mmproj-F16.gguf
       sha256: 7c5d39a9dc4645fc49a39a1c5a96157825af4d1c6e0961bed5d667a65b4b9572
       uri: huggingface://unsloth/Qwen3-VL-30B-A3B-Thinking-GGUF/mmproj-F16.gguf
+- name: runtimelabs-kayak-qwen3-vl-4b-q4
+  url: github:mudler/LocalAI/gallery/virtual.yaml@master
+  urls:
+    - https://huggingface.co/ayushadarsh7/RuntimeLabs-Kayak-Qwen3-VL-4B
+    - https://huggingface.co/mradermacher/RuntimeLabs-Kayak-Qwen3-VL-4B-GGUF
+  license: apache-2.0
+  description: |
+    RuntimeLabs Kayak is a 4B Qwen3-VL fine-tune for screenshot grounding on Kayak flight-search interfaces.
+    It predicts a single action and a normalized bounding box from an image and instruction.
+    This Q4_K_M GGUF build includes the F16 vision projector and uses an 8,192-token context with llama.cpp.
+  tags:
+    - llm
+    - gguf
+    - cpu
+    - gpu
+    - qwen3-vl
+    - multimodal
+    - vision
+    - screenshot-grounding
+  variants:
+    - model: runtimelabs-kayak-qwen3-vl-4b-q8
+  overrides:
+    backend: llama-cpp
+    context_size: 8192
+    known_usecases:
+      - chat
+    known_input_modalities:
+      - text
+      - image
+    mmproj: mmproj/RuntimeLabs-Kayak-Qwen3-VL-4B.mmproj-f16.gguf
+    options:
+      - use_jinja:true
+    template:
+      use_tokenizer_template: true
+    parameters:
+      model: RuntimeLabs-Kayak-Qwen3-VL-4B.Q4_K_M.gguf
+      temperature: 0
+    mmap: true
+  files:
+    - filename: RuntimeLabs-Kayak-Qwen3-VL-4B.Q4_K_M.gguf
+      sha256: 0286db34a9417e41361da1650364708000d437c782355fa15ae9395f1782abdc
+      uri: https://huggingface.co/mradermacher/RuntimeLabs-Kayak-Qwen3-VL-4B-GGUF/resolve/4228a5ebf59c63b1f809fd553fdde19ede7ac104/RuntimeLabs-Kayak-Qwen3-VL-4B.Q4_K_M.gguf
+    - filename: mmproj/RuntimeLabs-Kayak-Qwen3-VL-4B.mmproj-f16.gguf
+      sha256: 519eeb6616f5d014af407f28bfaad6ce5930468fadadaaf0c3cbfff270b763ad
+      uri: https://huggingface.co/mradermacher/RuntimeLabs-Kayak-Qwen3-VL-4B-GGUF/resolve/4228a5ebf59c63b1f809fd553fdde19ede7ac104/RuntimeLabs-Kayak-Qwen3-VL-4B.mmproj-f16.gguf
+- name: runtimelabs-kayak-qwen3-vl-4b-q8
+  url: github:mudler/LocalAI/gallery/virtual.yaml@master
+  urls:
+    - https://huggingface.co/ayushadarsh7/RuntimeLabs-Kayak-Qwen3-VL-4B
+    - https://huggingface.co/mradermacher/RuntimeLabs-Kayak-Qwen3-VL-4B-GGUF
+  license: apache-2.0
+  description: |
+    RuntimeLabs Kayak is a 4B Qwen3-VL fine-tune for screenshot grounding on Kayak flight-search interfaces.
+    It predicts a single action and a normalized bounding box from an image and instruction.
+    This Q8_0 GGUF build includes the F16 vision projector and uses an 8,192-token context with llama.cpp.
+  tags:
+    - llm
+    - gguf
+    - cpu
+    - gpu
+    - qwen3-vl
+    - multimodal
+    - vision
+    - screenshot-grounding
+  overrides:
+    backend: llama-cpp
+    context_size: 8192
+    known_usecases:
+      - chat
+    known_input_modalities:
+      - text
+      - image
+    mmproj: mmproj/RuntimeLabs-Kayak-Qwen3-VL-4B.mmproj-f16.gguf
+    options:
+      - use_jinja:true
+    template:
+      use_tokenizer_template: true
+    parameters:
+      model: RuntimeLabs-Kayak-Qwen3-VL-4B.Q8_0.gguf
+      temperature: 0
+    mmap: true
+  files:
+    - filename: RuntimeLabs-Kayak-Qwen3-VL-4B.Q8_0.gguf
+      sha256: 1e0109e2a5b92f3e369937f35e354652b5c7b4fb0dac040a459556e6af07b036
+      uri: https://huggingface.co/mradermacher/RuntimeLabs-Kayak-Qwen3-VL-4B-GGUF/resolve/4228a5ebf59c63b1f809fd553fdde19ede7ac104/RuntimeLabs-Kayak-Qwen3-VL-4B.Q8_0.gguf
+    - filename: mmproj/RuntimeLabs-Kayak-Qwen3-VL-4B.mmproj-f16.gguf
+      sha256: 519eeb6616f5d014af407f28bfaad6ce5930468fadadaaf0c3cbfff270b763ad
+      uri: https://huggingface.co/mradermacher/RuntimeLabs-Kayak-Qwen3-VL-4B-GGUF/resolve/4228a5ebf59c63b1f809fd553fdde19ede7ac104/RuntimeLabs-Kayak-Qwen3-VL-4B.mmproj-f16.gguf
 - name: qwen3-vl-4b-instruct
   url: github:mudler/LocalAI/gallery/qwen3.yaml@master
   urls: