Xiaomi just open-sourced a coding agent that beats Claude Code at 200-step tasks. 7,600 stars in 48 hours.

Xiaomi released MiMo Code under MIT license on June 10.

It hit 7,600 GitHub stars in two days.

This isn't a wrapper. It's a fork of OpenCode with a persistent memory architecture that solves the exact problem every coding agent hits: context amnesia over long sessions.

The numbers.

SWE-bench Verified: 82% vs Claude Code's 79%.
SWE-bench Pro: 62% vs 55%.
Terminal Bench 2: 73% vs 69%.

But here's what the benchmarks don't show.

Xiaomi ran a double-blind A/B test with 576 developers across 474 private repos. Under 200 execution steps, the two systems split roughly 50/50. Past 200 steps, MiMo Code's win rate jumped above 65%.

The memory architecture is the differentiator.

Four layers: project memory (MEMORY.md), session checkpoints, scratch notes, and per-task progress logs. An independent "checkpoint-writer" subagent updates blueprints in real time while the primary agent builds. When context window approaches limits, the system rebuilds from structured checkpoints.

Dream and Distill. The /dream command reviews historical sessions every ~7 days, deduplicates, and compresses into long-term memory. /distill mines past sessions for repeated workflows and packages them into reusable skills.

The model is bundled free. MiMo-V2.5: 310B parameters, 15B active per inference, 1M token context window. Zero configuration required. Also supports DeepSeek, Kimi, GLM, and any OpenAI-compatible API.

This is Xiaomi's third coding agent entry. The playbook matches DeepSeek and Alibaba: release capable tooling under permissive licenses at a fraction of US lab pricing, then convert developer mindshare into ecosystem.

Audit your current coding agent setup. If you're hitting context decay past 50 steps, the memory architecture alone justifies an evaluation. The MIT license means you can inspect the entire stack before committing.

SOURCE: https://github.com/XiaomiMiMo/MiMo-Code
VERIFIED: VentureBeat, Gadgets Now, Xiaomi official blog, GitHub repository
SIGNAL: Scaffolding and harness engineering are becoming as important as raw model capability. The coding agent wars just got a new contender with a fundamentally different architecture.