Releases·Model Release
Xiaomi releases MiMo-V2.6: "Frontier intelligence, all the modalities, built in public."
Xiaomi released MiMo-V2.6, marketing it as multimodal 'frontier intelligence, all the modalities, built in public,' with a launch page at mimo.xiaomi.com/mimo-v2-6. The company says the model's total training cost was $3.5M. It also ships a live 'benchmaxxing' dashboard for tracking benchmark performance.
85 engagement2 sourcesTue, Sep 22, 2026, 07:56 AM
AI Overview
- Xiaomi reports MiMo-V2.6's total training cost was $3.5M.
- Xiaomi ships MiMo-V2.6 with a live dashboard tracking benchmark performance.
- Xiaomi hosts MiMo-V2.6 details at mimo.xiaomi.com/mimo-v2-6.
Related
Social·Community
Xiaomi's MiMo-V2.6: An Omni-Modal Model Family Built to Improve Through Large-Scale Reinforcement Learning
Xiaomi's MiMo-V2.6 explainer: omni-modal model family built for self-improvement via reinforcement learning
Xiaomi's MiMo-V2.6 explainer: omni-modal model family built for self-improvement via reinforcement learning
0 engagement1 sourcerssSat, Sep 26, 2026, 04:16 AM
Releases·Model Release
MiMo-V2.6: Scaling Reinforcement Learning Towards Self-Improvement
A technical report introduces the MiMo-V2.6 series, an omni-modal model family trained by scaling reinforcement learning compute along batch size, environment diversity, and grader compute. The team built mid-training on multimodal corpora, froze the MoE router, and added multi-layer defenses against reward hacking to keep large-scale RL stable. Training dynamics, RL environments, and the RL framework are open-sourced for reproduction.
A technical report introduces the MiMo-V2.6 series, an omni-modal model family trained by scaling reinforcement learning compute along batch size, environment diversity, and grader compute. The team built mid-training on multimodal corpora, froze the MoE router, and added multi-layer defenses against reward hacking to keep large-scale RL stable. Training dynamics, RL environments, and the RL framework are open-sourced for reproduction.
0 engagement1 sourcearxivThu, Oct 8, 2026, 01:41 PM
Social·Community
Xiaomi Publishes Live Training Dashboard for Mimo 2.6 at mimo.xiaomi.com
Community watches Xiaomi MiMo 2.6 live training dashboard
Community watches Xiaomi MiMo 2.6 live training dashboard
137 engagement1 sourceredditWed, Sep 16, 2026, 08:49 PM
Releases·Model Release
MiMo-V2.6 update: tool-call repetition, diagnosed & fixed.
Xiaomi's MiMo team diagnosed and fixed a tool-call repetition issue in the MiMo-V2.6 model series, which had caused repeated identical tool calls burning context and stalling tasks. The root cause was a reward blind spot in scaling RL, fixed via a lightweight repetition-specialized RL teacher merged at ~4% of a full retrain's cost.
Xiaomi's MiMo team diagnosed and fixed a tool-call repetition issue in the MiMo-V2.6 model series, which had caused repeated identical tool calls burning context and stalling tasks. The root cause was a reward blind spot in scaling RL, fixed via a lightweight repetition-specialized RL teacher merged at ~4% of a full retrain's cost.
Updated Mon, Sep 28, 2026, 07:11 AM — Xiaomi confirmed via X post that community-reported tool-call repetition was diagnosed and fixed with a targeted RL upda
6 engagement1 sourceredditMon, Sep 28, 2026, 07:11 AM
Releases·Model Release
Xiaomi quietly uploads MiMo-V2.5-DFlash weights to Hugging Face
Xiaomi has uploaded the official DFlash weights for MiMo-V2.5-DFlash to Hugging Face. The model has 300B+ parameters and runs at 8-10 tokens/second on 2x24GB cards with offloading; DFlash could double that speed. The MTP head was shared but does not work yet.
Xiaomi has uploaded the official DFlash weights for MiMo-V2.5-DFlash to Hugging Face. The model has 300B+ parameters and runs at 8-10 tokens/second on 2x24GB cards with offloading; DFlash could double that speed. The MTP head was shared but does not work yet.
7 engagement1 sourceredditSun, Jul 12, 2026, 07:11 AM