M3OPD-Warmup-SFT-Video-9B
Paper: arXiv:2611.23034 · Project page · Code
Compared recipe of the M3OPD-Reasoner paper. SFT warm-up for video: full-parameter SFT of Qwen3.5-9B on the video SFT data of M3OPD-SFT-1.35M (folder video). This run trained on two Qwen3.8-Flash-Next traces per video question, and M3OPD-SFT-1.35M keeps the first of them. GSPO from this model gives the warm-up teacher of this domain (M3OPD-Warmup-Teacher-*-9B). The main recipe uses no SFT.
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-Warmup-SFT-Video-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.