🛠️ MiMo-V2.6 update: tool-call repetition, diagnosed & fixed. After the MiMo-V2.6 series models launched, we noticed them sometimes repeati…

🛠️ MiMo-V2.6 update: tool-call repetition, diagnosed & fixed. After the MiMo-V2.6 series models launched, we noticed them sometimes repeating identical or highly similar tool calls — burning context and stalling tasks, especially in MiMo Desktop, MiMo Code and OpenCode. Root cause A "reward blind spot" in scaling RL: when rewards only track final-answer correctness, inefficient behaviors along the way go unnoticed — and get amplified as training scales. Concretely, our flooding penalty only kicked in at >32 tool calls per turn, so anything below that threshold went completely unpunished. The fix We trained a light-weight repetition-specialized RL teacher — just 12 steps, ~7k examples — and merged it into the main model via MOPD, at roughly 4% of the cost of a full mixRL retrain. Repetition dropped sharply across harnesses and context lengths, while benchmarks held steady. 📦 Updated models, open-sourced (MOPD suffix): https://huggingface.co/collections/XiaomiMiMo/mimo-v26 📝 Full postmortem: https://mimo.xiaomi.com/blog/mimo-v2-6-tool-call-repetition Huge thanks to our community for the patience and feedback 🙏 Updated models go live on our API platform Sep 25, 06:00 (UTC+8), names unchanged. And for MiMo Desktop users: everyone's remaining quota in the current window will be reset.

Related

• We are releasing our first quantized checkpoints for the Qwen3.5 series of models, co-designed jointly with our inference engine to achieve …
• We’re dropping two open source SLMs this week. 1. One of them matches SOTA accuracy at up to 93x smaller. 2. The other one beats a recent Op…
• Hugging Face is becoming the platform for agents to use and build AI. Now they can call 1M HF spaces to do everything the latest specialized…

Source: Clem Delangue (X) | 2026-09-27