M3OPD-MixSFT-9B
a multimodal Qwen3.5 fine-tune for reasoning tasks
1 min readupdated
This model serves as the baseline (reference version for comparison) from the M3OPD-Reasoner study, created via full-parameter SFT of Qwen3.5-9B. It was trained on mixed SFT (supervised fine-tuning) data covering visual STEM problems, image perception tasks, and video question answering. Because it retains the exact architecture, tokenizer, processor, and chat template of the base Qwen3.5-9B model, it works anywhere the original is supported. It is well-suited for researchers and developers working on multimodal models and visual reasoning across images and video.
Key features
- Built on the capable Qwen3.5-9B model with 9.41B parameters
- Trained on the mixed M3OPD-SFT-1.35M dataset plus additional video QA trajectories
- Fully compatible with the Qwen3.5 ecosystem (Transformers, vLLM, SGLang)
- Uses the native chat template with thinking enabled
What you can do
- Run it locally or deploy it as a service via vLLM to handle multimodal requests
- Use it as a starting point for research and a baseline prior to the MixRL stage
- Solve visual STEM problems and analyze images using the \boxed{} answer format
Example
Launching the model using a vLLM server:
vllm serve ngqtrung/M3OPD-MixSFT-9B --max-model-len 40960Read more
DetailsArchitecture: image-text-to-text based on Qwen3.5-9B
Training step: step 1456 (start of the MixRL baseline)
License: Apache 2.0










Comments