moondream

/

moondream3-preview

like 558

Follow

moondream 965

Summary
This PR updates moondream/moondream3-preview to the new BF16 codepath, adds region‑head LN, enables finetune adapters (LoRA), and fixes spatial‑ref handling so spatial refs can be provided as inputs without re-encoding during answer generation.

New weights

• Added modelv2-00001-of-00004.safetensors … modelv2-00004-of-00004.safetensors
• Updated model.safetensors.index.json to point to modelv2-* as the new default
• Legacy model-0000x-of-00004.safetensors are retained for hard‑coded URL compatibility

Region model update

• Region head now applies LN before coord/size decoders (matches the new weights and backend parity)

Finetune / LoRA support

• Adapters are resolved via finetune_id@step and fetched from the finetune endpoint
• API-style model strings are supported (prefix ignored; /<finetune_id>@<step> is parsed)
• Example request format (API):
  { "model": "moondream3-preview/01K...@80", "question": "...", "image_url": "..." }
• Example model usage:
  model.query(image, question, settings={"adapter": "01K...@80"})