Xiaomi open-sources MiMo-V2-Flash for fast, dirt-cheap coding

MiMo-V2-Flash is Xiaomi's newly open-sourced Mixture-of-Experts foundation model, packing 309B total parameters with just 15B active parameters per token. It is built specifically for heavy-duty coding, mathematical reasoning, and agentic workflows, while remaining fast enough for daily chat.

What it can do and benchmark results

Software engineering is where this model clearly shines. On both SWE-bench Verified and the Multilingual benchmark, MiMo-V2-Flash grabbed the number one spot among all open-source models, holding its own against top proprietary systems.

For complex reasoning, it ranks among the top two open-source models on the AIME 2025 math benchmark and GPQA-Diamond for scientific knowledge.

It also comes loaded with developer-first features. A hybrid thinking mode lets you toggle between instant replies and step-by-step reasoning. It can generate functional HTML interfaces with a single prompt and hooks directly into modern coding tools like Cursor, Cline, and Claude Code. With a 256k context window, it can stay coherent across hundreds of tool-use rounds without losing track of project state.

How it stays fast and cheap

The inference speed clocks in at 150 tokens per second. Pricing is just as aggressive, set at $0.10 per million input tokens and $0.30 per million output tokens, according to blended benchmark figures from Artificial Analysis.

To hit these throughput numbers, Xiaomi used a hybrid attention setup that combines Global Attention and Sliding Window Attention in a 1 to 5 ratio, using a tight 128-token window. This design keeps the KV cache footprint small and predictable without tanking reasoning depth.

For inference, the model deploys native Multi-Token Prediction for self-speculative decoding. Instead of generating tokens strictly one by one, a lightweight draft block produces candidate tokens that the base model validates in parallel, delivering a measured 2.0x to 2.6x real-world speedup.

Post-training relied on a technique Xiaomi calls Multi-Teacher Online Policy Distillation. By having the student model learn from multiple specialized teacher models using dense token-level rewards, they matched peak RL reasoning performance using less than one-fiftieth of traditional compute.

Where to run it

Xiaomi has made the weights available and opened access on their mimo.xiaomi.com platform. You can test it online or grab an API endpoint to plug into your IDE.

Keep in mind that while only 15B parameters fire per token, you still have to host a 309B parameter model in memory if you plan to self-host, meaning significant enterprise-grade hardware is still required for local setups.

Why it matters

For builders who lean heavily on AI coding scaffolds, MiMo-V2-Flash looks like a massive win. You get frontier-grade benchmark performance on real-world bug fixing at a price that makes high-iteration agentic coding painless.

It also proves that clever architecture choices like tight sliding windows and native speculative decoding can drag massive MoE models out of latency hell. If third-party testing backs up these numbers in production, open-source coding agents just took a serious leap forward.