Switch to Qwen3.5

[gvzdv]

New Qwen vision model is out, and it's (on paper) better than the 2.5 that we use now.

I would like to try it out, which means we'll need to set aside time to swap models on Pegasus and test the new one.
There are several model variations (thinking/non-thinking, various sizes). I am considering a non-thinking (for speed) non-quantized 8B model — let's see if we can fit it on our GPU, especially if we can deprecate YOLO. Alternatives would be an 8-bit 8B model or an FP16 4B model.

[jeffbl]

Always a new model! Minor comment: be careful with "deprecate" which would be removing from the system forever. In this case, someone might want to run YOLO in the future, so we can simply remove it from our docker-compose.yml in favor of the LLM version, and get that memory back. In fact, we should definitely do so once we're running the LLM version.

[jeffbl]

qwen3.5 is out, and is a unified text/vision model. Needs testing as discussed in #1182. Hopefully this can be tested and merged into production sooner rather than later.

Questions:

• right now, we're running qwen3.5:4b on mallory. qwen3.5:9b is also available, and context window should be adequate. Are the results better?
• In my testing, qwen3.5:4b finds lots more discrete objects than yolo did. I don't know how this might mess with the end-user experiences on the Monarch (@Samyuktha0104) or audio-only experiences (@M-A89). If problematic, we can adjust the object detection prompt, but this would require discussion. (e.g., I'm thinking about additional constraints like "only find object that take up more than 5% of the entire photo" or "if an object is part of a larger whole that is also detected as an object, only include the overall object, e.g., instead of separately finding a person and a hand, just include the person, and make sure the bounding box encompasses all body parts."
• It seems slow to me, but I'm not sure if I'm off base, since this is a different model running on a slower GPU than pegasus, and the context window is boosted in open-webui settings. Maybe we should turn that down?