Qwen Cup multimodal reasoning competition showcases teacher‑distillation SFT and small‑dose GRPO RL on Qwen3.5‑0.8B, achieving 3rd place in finals. The team uses reward‑eval verbatim alignment and extensive data engineering, publishing results on GitHub with 11 stars. This demonstrates practical RL fine‑tuning for multimodal LLMs.
Opening Kapyn…