深度拆解 MiMo-V2.6：1M 上下文只是表面，2.5 万条轨迹才是底牌
Ranking
Overall
82
Content
95
Popularity
N/A
No observed public metrics; popularity remains neutral/archived.
Merged summary
TL;DR - Xiaomi released and open-sourced MiMo-V2.6, a multimodal model family built for long-context agent workloads. Its main advance is a unified, asynchronous reinforcement-learning system that generates 25,088 agent trajectories per update and extracts richer training signals from them.
- MiMo-V2.6 Pro has 1.02T total/42B active parameters, while Flash has 309B total/15B active parameters; both support 1M-token context and text, image, video, and audio inputs.
- Pro combines sliding-window and global attention, sparse MoE routing, and speculative decoding to reduce the cost of generating long, tool-using trajectories.
- Fully asynchronous GRPO separates rollout generation, environment execution, grading, and model updates, avoiding synchronization delays across tasks with widely varying completion times.
- A unified RL policy spans coding, general agents, vision, and cybersecurity; GRS/GAR rank trajectory quality beyond binary success, while MOPD2 reuses intermediate states from teacher and demonstration trajectories.
Sources (1)
深度拆解 MiMo-V2.6：1M 上下文只是表面，2.5 万条轨迹才是底牌
Public signals
N/A
TL;DR - Xiaomi released and open-sourced MiMo-V2.6, a multimodal model family built for long-context agent workloads. Its main advance is a unified, asynchronous reinforcement-learning system that generates 25,088 agent trajectories per update and extracts richer training signals from them.
- MiMo-V2.6 Pro has 1.02T total/42B active parameters, while Flash has 309B total/15B active parameters; both support 1M-token context and text, image, video, and audio inputs.
- Pro combines sliding-window and global attention, sparse MoE routing, and speculative decoding to reduce the cost of generating long, tool-using trajectories.
- Fully asynchronous GRPO separates rollout generation, environment execution, grading, and model updates, avoiding synchronization delays across tasks with widely varying completion times.
- A unified RL policy spans coding, general agents, vision, and cybersecurity; GRS/GAR rank trajectory quality beyond binary success, while MOPD2 reuses intermediate states from teacher and demonstration trajectories.