Can it run on a phone?

310 billion parameters (even with 15 billion active) require servers. Smaller compressed versions may reach powerful devices in the future — but V2.5 is designed for cloud hosting.

Does MiMo genuinely compete with the big models?

On some benchmarks — yes, especially in token efficiency and agentic tasks. On general-use scale, ecosystem, and experience — OpenAI, Anthropic, and Google still hold a significant advantage. But the gap is narrowing at a notable pace.

Why does open-source under MIT matter?

MIT means any developer can download the weights and host the model locally or commercially without restrictions. This undermines the value of closed-source APIs since any company can build its product without paying per-token fees to Anthropic or OpenAI.

Is Xiaomi an AI company?

Xiaomi started with phones but is investing heavily in AI. MiMo V1 was experimental, V2 stronger, V2.5 now genuinely competitive. A pattern similar to DeepSeek and Qwen from Alibaba — Chinese companies catching up at remarkable speed.

Xiaomi Launches MiMo-V2.5 — Open-Weight AI Model With 310B Parameters and 1M Token Context Window That Rivals Claude and Gemini at Half the Cost

Xiaomi Jumps Into Open-Source AI — MiMo-V2.5 Puts Rivals on Notice

Xiaomi is the latest company to release an open-weight AI model — MiMo-V2.5 claims to be a "major step forward in agentic capability and multimodal understanding."

The April 28, 2026 launch includes two variants: the standard MiMo-V2.5 and the flagship MiMo-V2.5-Pro. Both are released under the permissive MIT license — allowing free commercial use, fine-tuning, and modification without additional authorization.

Technical Architecture

Standard MiMo-V2.5: 310 billion total parameters (15 billion active). Trained on 48 trillion tokens. 1 million token context window. Natively multimodal supporting text, image, and video.

Flagship MiMo-V2.5-Pro: 1.02 trillion total parameters (42 billion active). Trained on 27 trillion tokens. 1 million token context window. Hybrid MoE architecture with 6:1 local-to-global attention ratio.

Key architectural innovation: the hybrid attention mechanism reduces KV-cache storage by nearly seven times while maintaining performance — allowing the model to "skim" large amounts of context and apply dense attention only to the most relevant 15% of data.

Benchmark Results — MiMo vs Competitors

On ClawEval for daily agentic tasks: V2.5-Pro achieves 64% Pass³ using approximately 70,000 tokens per trajectory — roughly 40-60% fewer tokens than Claude Opus 4.6, Gemini 3.1 Pro, and GPT-5.4 to achieve comparable results.

On SWE-bench Pro for software engineering: V2.5-Pro scores 57.2%, surpassing Claude Opus 4.6 (53.4%) and within 0.5 points of GPT-5.4 (57.7%).

On Artificial Analysis Intelligence Index: V2.5-Pro scored 54, tying with Kimi K2.6 for first place among open-weight models — surpassing DeepSeek V4 Pro.

Real Agentic Achievements

V2.5-Pro built a complete SysY compiler in Rust from scratch — including lexer, parser, and RISC-V assembly backend — in 4.3 hours with 672 tool calls, achieving a perfect 233/233 score. A task that typically takes a computer science major several weeks.

It also created a full-featured video editor with 8,192 lines of code over 11.5 hours.

Developer Pricing

$1.00 per million input tokens. $3.00 per million output tokens. Plus 100 trillion free tokens for a limited time.

This is significantly cheaper than Claude Opus 4.6 and GPT-5.4 for the same performance level — effectively because of the high token efficiency.