1 hour ago
16
Xiaomi just dropped an AI coding agent that beats Anthropic’s Claude Code at the tasks developers care about most: the long, grueling, multi-step ones that actually resemble real work.
MiMo Code V0.1, launched on June 11 under an MIT open-source license, is built on the MiMo-V2.5 model family. Its headline achievement is outperforming Claude Code on tasks exceeding 200 steps or requiring more than 1,000 tool calls. And it does this while burning through significantly fewer tokens, which translates directly into lower costs for developers.
The numbers tell the story
On Terminal-Bench 2.0, a benchmark designed to test AI coding agents on complex, multi-step terminal tasks, MiMo Code scored up to 86.7%. Claude managed 65.4% on the same benchmark. That’s not a marginal improvement. That’s a 21-percentage-point gap.
The SWE-bench Pro results paint a similar picture. MiMo Code achieved 57.2%, outperforming Claude Opus variants on the benchmark that tests an AI’s ability to resolve real-world GitHub issues.
MiMo Code accomplished all of this while using 40-60% fewer tokens across various agentic scenarios, meaning cheaper API bills for anyone running it at scale.
The model powering this performance is the MiMo-V2.5-Pro, a mixture-of-experts (MoE) architecture packing approximately 1.02 trillion parameters with a 1-million-token context capacity. That massive context window is what enables the agent to maintain coherence across those 200+ step tasks where other tools tend to lose the thread of what they’re doing.
Why long-context coding matters
MiMo Code’s persistent memory feature allows the agent to maintain project context across multiple sessions, so developers can pick up where they left off without re-explaining everything.
The terminal-native design operates directly in the developer’s existing terminal environment, reducing friction and keeping developers in their natural workflow.
Xiaomi has also integrated support for Claude Code via API within the MiMo terminal itself, so developers who want to compare outputs or use Claude for specific subtasks can do so without switching tools.
Xiaomi’s quiet AI buildout
Xiaomi has been steadily building its AI capabilities, with the MiMo V2-Flash model arriving in late 2025 as an earlier entry in the series. By April 2026, daily token usage across the MiMo platform exceeded 1 trillion tokens.
The MIT license means MiMo Code can be downloaded, modified, and deployed by anyone for any purpose, at no cost.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.
English (US) ·