Breaking

Xiaomi Releases Open-Weight MiMo-V2.6 Pro and Flash Models for Agentic AI Workloads

September 21, 2026 at 17:56 EDT

Foundation Models
AI Agents
Open Source

Xiaomi has released MiMo-V2.6-Pro-RL and MiMo-V2.6-Flash-RL as open-weight models, moving its MiMo family deeper into agentic coding, tool use and multimodal long-context work. The release also includes a MiMo-V2.6-Distill-Qwen-9B checkpoint and research assets intended to support further agentic reinforcement-learning experiments.

Sept. 21, 2026 · Xiaomi

MiMo-V2.6 moves into the open-weight front tier

Xiaomi released Pro and Flash checkpoints with agentic RL training assets, a 1M-token context window, multimodal inputs and MIT-licensed weights.

Intelligence Index

Cost per task

$0.13

Output speed

134.3

tokens/s

Context window

tokens

Active parameters: Pro vs Flash

The larger Pro checkpoint activates 42B parameters per token, compared with 15B for Flash.

Training story

Mixed RL

coding, agents, visual work, security

→

Groupwise grading

rank rollouts beyond pass or fail

→

Distillation

extend behavior to harder tasks

Why developers care

MIT weights, 1M context, multimodal inputs and strong agentic benchmarks make the release attractive for coding and tool-use systems.

Where caution remains

The models are hardware-heavy, Pro is verbose in evaluation, and results remain uneven on some security and coding benchmarks.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.