Xiaomi has released and open-sourced MiMo-V2.5-Pro, a model the company describes as its most capable to date, with stronger agentic behavior, longer-horizon coherence, and improved instruction following compared with MiMo-V2-Pro.
According to Xiaomi, MiMo-V2.5-Pro is a "1.02T-parameter Mixture-of-Experts" model with "42B active parameters," built on a hybrid-attention architecture and a "1M-token context window." The company also claims that, when paired with an appropriate harness, the model can sustain tasks spanning "more than a thousand tool calls." Xiaomi says the model is now rolled out across its API Platform, AI Studio, and other surfaces, with pricing unchanged and the model tag set as `mimo-v2.5-pro`.
The company highlights several internal demonstrations of the model’s capabilities, though the results should be treated as vendor-reported. In one case, Xiaomi says MiMo-V2.5-Pro completed a complete SysY compiler in Rust — including lexer, parser, AST, Koopa IR codegen, a RISC-V assembly backend, and performance optimization — in 4.3 hours across 672 tool calls, passing 233/233 hidden tests. Xiaomi also points to a desktop video editor it claims the model produced over 11.5 hours and 1,868 tool calls, with features including a multi-track timeline, clip trimming, cross-fades, audio mixing, and export support.
Another example centers on analog EDA. Xiaomi says MiMo-V2.5-Pro was wired into an ngspice simulation loop with Claude Code as the harness and, after about an hour of iteration, produced an FVF-LDO design that met all target metrics. The company further claims the model showed "harness awareness," describing behavior that made use of the surrounding environment to manage memory and context toward the final objective.
Xiaomi also says it scaled post-training compute to improve coding performance. Its in-house MiMo Coding Bench is presented as a test suite for repo understanding, project building, code review, structured artifact generation, planning, SWE, and related tasks within agentic frameworks such as Claude Code. The company states that MiMo-V2.5-Pro improves real-world coding scenarios and can be integrated into scaffolds such as Claude Code, OpenCode, and Kilo.
Token efficiency is another area Xiaomi emphasizes. On ClawEval, the company says MiMo-V2.5-Pro reached 64% Pass^3 using around 70,000 tokens per trajectory, which it characterizes as roughly "40–60% fewer tokens" than Claude Opus 4.6, Gemini 3.1 Pro, and GPT-5.4 at comparable capability levels.
Alongside the model release, Xiaomi says it updated its inference infrastructure and Token Plan, including a reset of used Credit balances for users who bought a Token Plan before 14:00 UTC on April 21. The company also says MiMo-V2.5-Pro is fully open-sourced under a permissive license, with weights, tokenizer, and model card available on Hugging Face. The public specs list MiMo-V2.5-Pro-Base with a 256K context window and MiMo-V2.5-Pro with a 1M-token context window, both using FP8 (E4M3) mixed precision.
Source: Xiaomi MiMo