M3OPD-MixSFT-9B

Paper: arXiv:2611.23034 · Project page · Code

MixSFT baseline of the M3OPD-Reasoner paper: full-parameter SFT of Qwen3.5-9B on the mixed SFT data of all three domains, namely the image STEM and image perception self traces of M3OPD-SFT-1.35M and 311,396 Qwen3.5-9B self traces of the video questions, which are not in M3OPD-SFT-1.35M. It is the start of the MixRL baseline (M3OPD-MixRL-Baseline-9B). It is a baseline, not the main model.

All checkpoints of the recipe are in the M3OPD-Reasoner collection. The data are M3OPD-SFT-1.35M and M3OPD-RL-193K.

Usage

The model has the architecture, tokenizer, processor and chat template of Qwen3.5-9B, so it runs wherever Qwen3.5-9B runs (Transformers, vLLM, SGLang). We train and evaluate with the native Qwen3.5 chat template, thinking enabled, no system prompt, and one answer-format line after the question, for example Put your final answer within \boxed{}.

vllm serve ngqtrung/M3OPD-MixSFT-9B --max-model-len 40960

Citation

@article{m3opdreasoner2026,
  title   = {M$^3$OPD-Reasoner: An Open Recipe for Post-Training Multimodal Reasoning
             with Multi-Teacher On-Policy Distillation},
  author  = {Nguyen Quang Trung and Hoang Phan and Shuo Sun and Longyin Zhang and AiTi Aw},
  journal = {arXiv preprint arXiv:2611.23034},
  year    = {2026}
}

License

Apache 2.0, the license of Qwen3.5-9B.

Downloads last month
28
Safetensors
Model size
9B params
Tensor type
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for ngqtrung/M3OPD-MixSFT-9B

Finetuned
Qwen/Qwen3.5-9B
Finetuned
(1,427)
this model

Datasets used to train ngqtrung/M3OPD-MixSFT-9B

Collection including ngqtrung/M3OPD-MixSFT-9B

Paper for ngqtrung/M3OPD-MixSFT-9B