MosaicAgent-8B

MosaicAgent-8B is a vision-language agent trained to use the Mosaic image workspace, based on Qwen3-VL-8B-Instruct.

Code · MosaicBench · Paper

Format and provenance

This checkpoint uses Qwen3VLForConditionalGeneration. The weights contain 750 tensors in four safetensors shards, stored in BF16. Tokenizer, image/video processor configuration and the chat template accompany the weights. These are fine-tuned weights, not the unchanged upstream model.

The experimental archive identifies this checkpoint as the image-workspace agent. Exact training data release and paper-result mapping are tracked separately. MosaicBench is evaluation data; it is not presented as this model's training dataset. No unverified benchmark scores are included here.

Run with Mosaic

Install the Mosaic code and prepare a dataset as described there. Serve this checkpoint with a Qwen3-VL-compatible vLLM installation:

vllm serve gengyuanmax/MosaicAgent-8B --dtype bfloat16 --port 8000
mosaic-run --dataset mosaicbench --env tool \
  --model-name gengyuanmax/MosaicAgent-8B --endpoint http://localhost:8000/v1

Set context length and image limits for your hardware and workload. The explicit BF16 setting is intentional: the original configuration declares float32 while the stored tensors are BF16.

Use the Mosaic runner to execute tools and return generated images to the model. A plain text generation call alone does not execute the image workspace.

For a direct two-image load/generation check, use verify_multimage.py with two local images. For MosaicAgent, this is only a loading check, not a tool-use evaluation.

Validation status and limitations

The archive index, shard boundaries, tensor inventory and accompanying metadata have been checked. Both checkpoints passed BF16 loading and two-image generation on an NVIDIA RTX A6000 with Transformers 4.57.1 and PyTorch 2.6.0. MosaicAgent also completed a crop tool call, received the generated image and returned a final answer. These are limited functional checks, not a benchmark accuracy evaluation. CPU protocol tests of the Mosaic runner use a simulated endpoint and do not demonstrate this checkpoint's quality.

The model can produce incorrect answers or tool arguments. It has not been validated for medical, legal or other high-stakes decisions. Third-party datasets retain their own terms.

License and attribution

The upstream Qwen model declares Apache License 2.0. The included LICENSE reproduces that license. Credit for the base model belongs to the Qwen team; this release contains Mosaic fine-tuning modifications. Dataset licenses do not change under this model license.