Xiaomi's MiMo-V2.5-Pro Matches Claude Opus on Coding Benchmarks and Uses Half the Tokens

Xiaomi released MiMo-V2.5-Pro on April 27, 2026. It’s an open-source coding model built on a Mixture-of-Experts architecture with 1.02 trillion total parameters and 42 billion active parameters per request. Weights, tokenizer, and documentation are all on Hugging Face under a permissive license.

The benchmark numbers are what got attention. On SWE-bench Verified, MiMo-V2.5-Pro scores 78.9. On SWE-bench Pro, 57.2, which puts it in the same range as Claude Opus 4.6. On Terminal-Bench 2.0, a benchmark specifically designed for long-horizon terminal tasks, it scores 68.4.

The more interesting claim is efficiency. On ClawEval, the agentic benchmark Xiaomi built to measure multi-step coding tasks, MiMo-V2.5-Pro hits 64% Pass³ while using roughly 70,000 tokens per trajectory. Comparable results from Claude Opus 4.6, Gemini 3.1 Pro, and GPT-5.4 cost 40 to 60 percent more tokens to achieve similar scores.

That gap matters for how much it actually costs to run the model on long tasks.

Architecture

MiMo-V2.5-Pro is a Mixture-of-Experts model, meaning the full 1.02 trillion parameters don’t all fire on every request. Each request activates 42 billion parameters, keeping inference costs closer to a 42B dense model than to a 1T one. The context window is 1 million tokens. Training used FP8 mixed precision on 27 trillion tokens.

The model supports text and code. It does not appear to have the multimodal capabilities of the earlier MiMo-V2 series.

What It Can Do in Practice

Xiaomi published two demonstration tasks in the launch materials.

First: the model was given a SysY-to-RISC-V compiler project with a hidden test suite. It worked autonomously for 4.3 hours, made 672 tool calls, and passed 233 out of 233 tests. The entire workflow ran without human intervention.

Second: it built a desktop video editor from scratch in 11.5 hours, writing 8,192 lines of code across 1,868 tool calls. The final application included a multi-track timeline, audio tools, and an AI voice-over feature.

Both of these are the kind of task that requires sustained context and coherent multi-step planning, not just single-shot code generation. The token efficiency numbers make more sense in this context: a model that needs fewer tokens to complete a multi-hour agentic task isn’t just cheaper to run, it can sustain longer trajectories before hitting context limits.

Where It Sits in the Open-Source Landscape

The open-source coding model space has gotten crowded. Kimi K2.5, DeepSeek, Qwen 3.5, and others are all competing on similar territory: large-parameter MoE models that approach frontier performance at a fraction of the cost.

What’s different about MiMo-V2.5-Pro is the explicit focus on agent efficiency. Most benchmarks measure capability (can the model do the task?). Xiaomi’s ClawEval numbers measure efficiency too (how many tokens does it take to do the task?). That’s a more practical metric for anyone running agents in production, where token costs compound across thousands of runs.

Access

MiMo-V2.5-Pro is available in three ways:

• Xiaomi’s interactive studio — no setup required
• Xiaomi’s API platform — change your model identifier to mimo-v2.5-pro
• Hugging Face — full open-source weights for local deployment

The Hugging Face release includes the tokenizer and configuration files needed to run inference locally.

---

Sources: MiMo-V2.5-Pro official page, The Decoder, VentureBeat