મીમો v2.5 પ્રો

Xiaomi's flagship open Mixture-of-Experts model, at 1.02T total and 42B active parameters, built for the most demanding agentic, software-engineering and long-horizon work. Xiaomi describes it as a leap in agentic coherence, sustaining complex trajectories across thousands of tool calls with strong instruction following over a million-token context. Its showcase runs include writing a compiler from scratch across hundreds of tool calls.

Xiaomi does not argue this model with a benchmark table first. It argues it with three jobs it left running. The first is a compiler: a university course project asking for a complete SysY compiler in Rust from scratch - lexer, parser, syntax tree, intermediate representation, RISC-V backend and performance work - which the model finished in 4.3 hours across 672 tool calls, passing all 233 hidden tests. The lab's reading of the trace is the interesting part: the first compile already passed 137 of 233, which suggests the architecture was designed correctly before a single test ran, and when a refactor at turn 512 regressed two cases the model diagnosed and recovered rather than thrashing. The second job produced a desktop video editor with a multi-track timeline, trimming, cross-fades, audio mixing and export - 8,192 lines over 1,868 tool calls and 11.5 hours. The third wired the model into a circuit simulator to design an analog voltage regulator in a 180nm process, landing six specifications simultaneously in about an hour of closed-loop iteration.

The behaviour Xiaomi names for all three is harness awareness: the model makes full use of what its environment offers, manages its own memory, and shapes how its context gets populated toward the goal. That is a different claim from a higher score, and it is the one the lab leads with - along with reliable adherence to subtle requirements buried in context and coherence held across very long inputs.

The efficiency argument is quantified. On its daily-agent benchmark Xiaomi reports 63.8 at roughly 70,000 tokens per trajectory, which it puts at 40 to 60% fewer tokens than the frontier closed models it compares against at comparable capability. Elsewhere on its published table: 72.9 on the multi-domain agent suite, 57.2 on SWE-Bench Pro, 78.9 on SWE-bench Verified, 68.4 on Terminal-Bench 2.0, and 34.0 on Humanity's Last Exam without tools rising to 48.0 with them.

The architecture is built for that token budget. Local sliding-window and global attention interleave at six to one with a 128-token window, which Xiaomi says cuts key-value cache storage by nearly seven times at long context while a learnable attention-sink bias preserves quality; a lightweight multi-token prediction module roughly triples output throughput and speeds up reinforcement-learning rollouts. Pre-training ran on 27 trillion tokens in FP8 mixed precision at a native 32K length before the window was extended. Where it stops is stated by the lab's own charts rather than in prose: it labels its internal coding result as closing the gap to a leading closed model rather than passing it, and on the hardest-exam and implementation-ranking rows the closed frontier is still ahead.