Xiaomi open‑source model excels coding

Xiaomi didn’t just ship another chatbot model. It open-sourced MiMo-V2.5-Pro on April 27, and the real pitch is software work — coding agents, long tool chains, and tasks that look more like “finish the project” than “answer the prompt.” That matters because open models still tend to fall apart when a job gets messy, long, or stateful. Xiaomi is claiming this one holds together much better, and the release gives developers actual weights to run and fine-tune, not just an API. (mimo.xiaomi.com) ### What actually shipped? MiMo-V2.5-Pro is Xiaomi’s new flagship MiMo model. It’s a mixture-of-experts system with 1.02 trillion total parameters, 42 billion active parameters, and a 1 million-token context window. Xiaomi says it has fully open-sourced the model under a permissive license, with weights, tokenizer, and model card available through Hugging Face, while also rolling it out on Xiaomi’s own API platform and studio tools. (mimo.xiaomi.com) ### Why are people calling it a coding model? Because Xiaomi is positioning it around software engineering and agent tasks, not just general chat. The company’s own benchmark table puts MiMo-V2.5-Pro at 57.2 on its “Coding Agent” eval and 73.7 on SWE-Bench Pro. It also highlights a concrete demo: the model built a SysY compiler in Rust from scratch — lexer, parser, IR generation, RISC-V backend, and op(mimo.xiaomi.com) with a perfect 233/233 on the hidden test suite. That’s the sort of example that gets developers’ attention fast. (mimo.xiaomi.com) ### What makes the agent angle different? A coding assistant can write a file. A coding agent has to keep going after the first file — inspect outputs, call tools, recover from mistakes, and stay coherent across a long chain of steps. Xiaomi’s claim is that V2.5-Pro can handle more than a thousand tool calls when the harness is set up properly, and that it follows subtle instructions better over very l(mimo.xiaomi.com)” but “fewer collapses in the middle of a real workflow.” (mimo.xiaomi.com) ### Is this brand new from Xiaomi? Not exactly. Xiaomi has been building toward this for a few months. MiMo-V2-Flash launched in December 2025 as an open model focused on reasoning, coding, and agentic scenarios, with a 256k context window and explicit support for tools like Claude Code, Cursor, and Cline. Before that, MiMo-V2-Pro arrived on March 18, 2026, with stronger agent benchmarks and a 1M-token(mimo.xiaomi.com)up. (mimo.xiaomi.com) ### Why do developers care about open weights here? Because open weights change the economics and the control surface. If a team wants a model inside an internal dev tool, browser agent, code-review bot, or game-generation stack, running a model they can inspect, adapt, and host is a very different proposition from renting a black-box API. Xiaomi’s Hugging Face page shows strong early traction too — Mi(mimo.xiaomi.com)d the Pro model had already drawn thousands more. (huggingface.co) ### Is Xiaomi really competitive here? On Xiaomi’s own published comparisons, yes — at least enough to be taken seriously. V2.5-Pro sits close to Claude Opus 4.6 on the company’s coding-agent metric and ahead of Gemini 3.1 Pro there, though still below GPT-5.4 in that same table. Earlier MiMo-V2-Flash also claimed the top open-source spot on SWE-Bench Verified and Multilingual. The catch is that these are v(huggingface.co)e whether outside developers can reproduce the same behavior in real repos and tool chains. (mimo.xiaomi.com) ### What’s the bigger picture? Xiaomi is turning into a real open-model player, not just a hardware brand dabbling in AI. Its GitHub and Hugging Face orgs now span language, audio, vision, embodied, and agent tooling projects, including MiMo-Skills for agent capabilities. That matters because the open-model race is shifting from “who has a decent base model” to “who has a model developers can actually build systems around.” (github.com) ### Bottom line? The news isn’t just that Xiaomi released a model. It’s that Xiaomi released an open model aimed squarely at coding agents — with huge context, big tool-use claims, and enough benchmark weight to make engineers try it. If the outside results hold up, MiMo-V2.5-Pro could become one of the more important open options in the developer stack this year. (mimo.xiaomi.com)