Xiaomi releases trillion-parameter MiMo-V2-Pro agent model
Xiaomi has put MiMo-V2-Pro, a 1T-parameter model with a 1M-token context window built for agent workflows, on a paid API and claims performance near Claude Opus 4.6 at about one-fifth the price.
Xiaomi announced MiMo-V2-Pro, which it calls its flagship foundation model for agent work. The model has more than 1T total parameters and 42B active parameters. Xiaomi says that is about three times the size of the earlier MiMo-V2-Flash. It supports a context window of up to 1M tokens.
The model keeps the hybrid attention design introduced in V2-Flash and raises the hybrid ratio from 5:1 to 7:1, which Xiaomi says keeps inference efficient at the larger size. A lightweight multi-token prediction (MTP) layer is used to speed up generation. For post-training, Xiaomi used SFT and RL across a range of agent scaffolds, aimed at tool use and multi-step reasoning.
Xiaomi says the model ranks eighth worldwide and second in China on Artificial Analysis's intelligence index. It also says the model ranks near the top of PinchBench and ClawEval, the benchmarks for the open-source OpenClaw agent framework. Most of its comparisons with Claude come from its own testing: Xiaomi says that in hands-on use the model surpasses Claude Sonnet 4.6 and approaches Opus 4.6, and that internal engineers rated its coding close to Opus 4.6. The company also places it in the same tier as Claude 4.5 Sonnet, GPT-5.2 and Gemini 3.0 Pro on coding-agent and tool-use tasks. It did not publish independent figures to support these claims. Xiaomi says "Hunter Alpha," a model tested anonymously before launch, was an early version of V2-Pro.
API pricing is $1 per million input tokens and $3 per million output tokens up to 256K context. From 256K to 1M tokens, it rises to $2 and $6.
The page's headline refers to a "MiMo-V2.6 series" of three new models, but its content covers only V2-Pro. No details on any V2.6 models were provided.