What Xiaomi Actually Released#
Xiaomi's MiMo AI team has open-sourced MiMo Code V0.1.0, a terminal-native AI coding assistant targeting the same agentic, multi-file coding workflows that Anthropic's Claude Code occupies. The release went public on or around June 11, 2026. Unlike Claude Code, which runs as a closed commercial product, MiMo Code is open-sourced: a direct play for developer trust and community adoption in a market where auditability increasingly matters.
The tool is positioned not as a chat-style autocomplete but as a multi-step agent: it plans, executes shell commands, edits files, and iterates across long task horizons. That framing is the core of Xiaomi's benchmark claim.
The Benchmark Claim and Its Limits#
That distinction matters. Vendor-run evaluations against competitors have a structural incentive to frame the task distribution favorably. A vendor-run A/B beta scored by its own participants is real-world signal, and a stronger one than a satisfaction survey, but it is not an independent benchmark. Xiaomi does report against a named public suite, SWE-Bench Pro, where it says its experimental Max Mode gains 10 to 20 percent over single sampling at roughly four to five times the token consumption. That is MiMo Code measured against itself, though, and the headline "beats Claude Code" framing still rests on the internal human-in-the-loop comparison rather than an independent third-party run.
Treat the headline performance claim as directionally interesting but not yet independently validated.
Where MiMo Code's long-horizon advantage is plausible on its face: managing context coherently across 200+ sequential steps is genuinely hard, and it's a known weak point for models with shorter effective context windows or poor state management. Xiaomi attributes its long-horizon edge to architecture-level design (checkpointing, layered memory, and persistent project knowledge) rather than thin prompt scaffolding.
Open Source and the Free API Window#
Two things differentiate MiMo Code from most commercial peers right now.
Xiaomi's own repository documentation describes MiMo Code as a terminal-native assistant that reads and writes code, runs shell commands, manages Git, and keeps a persistent memory of the project across sessions, which is close to how Claude Code operates. The terminal-native approach avoids IDE lock-in and fits into CI/CD pipelines more naturally than GUI-first tools.
Who This Is Actually For#
MiMo Code is most credibly interesting to three audiences.
It is less relevant for teams who need proven, independently benchmarked performance today, or whose primary use case is short interactive coding sessions where the long-horizon advantage is moot.
The comparison to Claude Code is the right one to make, since Claude Code is the incumbent terminal-native agent. But until an independent evaluation replicates Xiaomi's results on a named public benchmark with a stated date, the "beats Claude Code" framing should be held loosely.
What to Watch#
The immediate question is independent replication: does MiMo Code's long-horizon advantage hold on SWE-bench Verified or another public agentic coding benchmark run by a party without a stake in the outcome? Watch also for the post-promotion API pricing, which will determine whether the cost-per-task math actually competes with Claude Code at scale. If the open-source community picks up the repo and starts stress-testing it, that signal will arrive faster than any official benchmark.
Frequently asked questions
Is MiMo Code free to use?
As of launch, Xiaomi is offering limited-time free access through its built-in MiMo Auto channel, which it says runs on MiMo-V2.5 with a 1-million-token context. Exact duration and post-promotion pricing have not been published.
What makes a 200-step agentic task different from normal AI coding assistance?
A 200-step task requires the agent to maintain coherent state, plan ahead, and execute dozens of sequential actions (file edits, shell commands, test runs) without losing context. Most coding copilots are optimized for single-shot or short-session interactions, not sustained autonomous workflows of this length.
Can MiMo Code be self-hosted?
Yes, and the licensing is worth reading before you commit. The source is MIT, but the repository applies a separate Use Restrictions document on top of it, and Xiaomi-hosted model services carry their own terms. Self-hosting means teams can run the agent on their own infrastructure rather than routing code through a third-party API.
Discussion