Xiaomi MiMo Code claims edge over Claude Code on 200-step tasks

The endurance problem - agents degrading as tasks stretch past dozens of steps - is the current frontier of practical coding-agent engineering. MiMo Code's design answers it with engineering rather than model scale: a dedicated subagent owns record-keeping, checkpointing state and summarizing context before the window fills, so the main agent never starts from scratch. That pattern is portable to any agent framework, which makes this release worth attention even from teams that will never run a Xiaomi model.

Xiaomi released and open-sourced MiMo Code V0.1 under an MIT license, a terminal-native AI coding agent built as a fork of the open-source OpenCode project. Launch coverage from The New Stack and VentureBeat dates the release to June 10-11, 2026; Xiaomi's official announcement page describes the release and its architecture. The agent installs via a single terminal command or npm on Windows, ships with Xiaomi's MiMo-V2.5 model built in (free for a limited time, and described by Xiaomi as comparable to Claude Sonnet 4.6), and supports third-party models including DeepSeek, Kimi, and GLM.

Xiaomi's official announcement reports a controlled experiment in which MiMo Code and Claude Code ran the exact same MiMo model, isolating the harness itself: MiMo Code scored 73% vs Claude Code's 68% on Terminal-Bench 2, and 62% vs 57% on SWE-Bench Pro. Some launch coverage circulated larger figures (86.7% vs 65.4% on Terminal-Bench 2.0, and 57.2% on SWE-bench Pro) drawn from different configurations; accounts differ, and the official controlled-experiment numbers are the conservative, like-for-like comparison. Separately, Xiaomi cites an internal beta with a human A/B evaluation across 576 developers working in real private repositories, where MiMo Code's win rate against Claude Code rose above 65% once tasks passed 200 execution steps. Every figure here is vendor-run; none has been independently replicated yet.

The memory system uses layered persistence - project memory, conversation checkpoints, and per-task progress logs - and VentureBeat's teardown documents an implementation built on SQLite FTS5 full-text search, including a scratch-notes layer. An independent subagent saves state automatically and produces a clean summary when the context window nears capacity, so the main agent continues instead of restarting. A /dream command runs every seven days, merging, deduplicating, and compressing accumulated memory into a compact current state. A Compose mode (triggered with the Tab key) runs a full design-plan-code-test-review loop from a single instruction.

Harness-versus-harness competition is displacing model-versus-model comparison in coding agents; GitHub's recent Copilot token-efficiency work points the same direction. An MIT-licensed, documented memory stack lowers the cost of experimenting with these ideas, and OpenCode's permissive base makes vertical forks straightforward. The obvious caveats: vendor benchmarks routinely flatter the vendor, Terminal-Bench and SWE-Bench Pro configurations materially affect scores, and long-horizon win rates from an internal beta are not a public benchmark.

• •Independent Terminal-Bench 2 runs pitting MiMo Code against Claude Code and Codex CLI under matched configurations
• •Community adoption of the memory-subagent pattern in other OpenCode forks and agent frameworks
• •Whether Xiaomi publishes the evaluation harness behind its 576-developer A/B study