← Back to feed

Xiaomi’s MiMo AI team released MiMo Code V0.1.0 on June 10, an open-source terminal coding agent licensed under MIT. The release claims Claude Code loses to MiMo Code on long-horizon tasks exceeding 200 execution steps — a benchmark Anthropic’s agent is not specifically designed for, and one that Xiaomi says matters more as agents take on full software projects rather than individual file edits.

What It Is

MiMo Code is a fork of OpenCode, the open-source terminal coding agent, extended with four proprietary subsystems:

• Cross-session memory: A persistent memory layer that retains project understanding across terminal restarts. The agent builds an evolving model of the codebase that doesn’t reset at session end.
• Intelligent context management: Dynamic pruning of what goes into the context window, with automatic decisions about what to include from project history.
• Subagent orchestration: The ability to spawn and coordinate multiple parallel coding subagents for different parts of a task.
• Self-improvement via dream/distill: An asynchronous post-session phase where the agent reviews its own decisions and updates its internal heuristics — a lightweight form of online RL from its own trajectories.

It installs in one command: curl -fsSL https://mimo.xiaomi.com/install | bash. Windows users can use npm install -g @mimo-ai/cli.

The Model Angle

On first launch, MiMo Code prompts users to pick a model access method. The default is MiMo Auto — free access, for a limited time, to MiMo-V2.5. That model has 310 billion total parameters with 15 billion active per inference, a 1-million-token context window, and native multimodal input. Xiaomi positions it as competitive with Claude Sonnet 4.6 on agentic multimodal tasks. Users can alternatively import from an existing Claude Code configuration or bring their own API key for any supported provider.

The “free for a limited time” framing is significant: routes all context through Xiaomi’s servers. For individual developers, that’s a zero-cost entry point into a frontier agent. For enterprises with IP or data-residency policies, it is a blocker.

Benchmark Claims

The benchmark edge Xiaomi claims over Claude Code is self-reported, based on an internal beta survey of 576 developers, and the specific focus on 200-step tasks is a metric Anthropic doesn’t publish. That framing is deliberate: Claude Code is optimised for typical single-session coding tasks. MiMo Code is being positioned as the tool for runs that go hundreds of steps deep — debugging complex distributed systems, multi-file refactors, or long agentic research loops.

Whether the claim holds up on independent benchmarks is unresolved. The V0.1.0 version number is honest signalling about maturity.

Strategic Context

Xiaomi’s MiMo division has moved quickly. MiMo-7B shipped in April 2025 as a reasoning model. MiMo-VL followed with vision-language capability. MiMo-V2-Pro in March 2026 reached 1 trillion parameters. MiMo-V2.5 launched in April 2026 as the multimodal flagship.

MiMo Code is the first consumer-facing product that puts all of that together as a developer tool. The OpenCode fork strategy means Xiaomi inherits an existing ecosystem — multiple providers, TUI interface, LSP support, MCP, plugins — and ships differentiation on top rather than from scratch. The MIT license removes any friction for forks or contributions.

For the coding agent market, another capable open-source entrant changes the floor. Claude Code, Cursor, and GitHub Copilot’s CLI all compete on proprietary scaffolding over closed models. MiMo Code offers a route to frontier-class long-horizon agent capability with full local control over model selection.

The 5,300 GitHub stars inside 48 hours suggests developer appetite for the value proposition.