Xiaomi Open-Sources MiMo Code AI Coding Agent with Cross-Session Memory

Xiaomi enters the AI coding agent race

Xiaomi has released MiMo Code V0.1.0, an open-source terminal-native AI coding assistant that the company claims outperforms Anthropic's Claude Code on extended multi-step development tasks. The tool is available now on GitHub under an MIT license and installs with a single command on macOS, Linux, and Windows.

Announced June 10, 2026, MiMo Code is a fork of the open-source OpenCode agent that Xiaomi has extended with a novel memory architecture designed to solve a persistent problem in AI coding workflows: context degradation over long sessions. The company is bundling limited-time free access to MiMo-V2.5, its multimodal flagship model with a million-token context window, requiring no registration.

Why it matters

AI coding agents typically lose track of earlier decisions and project conventions as their context windows fill, forcing developers to repeatedly re-explain their work. Xiaomi's approach—storing project state in persistent structures rather than relying on context compression—addresses a real pain point that affects productivity in extended development sessions. If the architecture proves robust, it could shift how enterprises evaluate coding agent platforms, particularly for complex, multi-day projects where session continuity matters more than raw single-task performance.

Cross-session memory replaces context compression

MiMo Code attacks context degradation with a four-layer memory system powered by SQLite full-text search: a persistent MEMORY.md project file, session checkpoints, scratch notes, and per-task progress logs. Rather than forcing the primary coding agent to pause and take notes, the system deploys an independent "checkpoint-writer" subagent that updates structured records in real time.

When the context window approaches capacity, the system rebuilds the environment from these checkpoints with relevant context intact. Two self-improvement mechanisms round out the design: a /dream command that reviews and compresses historical sessions roughly every seven days, and a distill function that mines past work for repeated workflows that can be automated.

Benchmark claims and caveats

According to figures Xiaomi published, MiMo Code paired with MiMo-V2.5-Pro scored 82 percent on SWE-bench Verified versus 79 percent for Claude Code with Claude Sonnet 4.6, 62 percent versus 55 percent on SWE-bench Pro, and 73 percent versus 69 percent on Terminal Bench 2.

The harness itself accounts for roughly five percentage points of the gain. Running the same MiMo model in both harnesses, MiMo Code scored 62 percent on SWE-bench Pro versus 57 percent for Claude Code, and 73 percent on Terminal Bench 2 versus 68 percent—improvements attributable to the agent system rather than the underlying model.

Xiaomi notably compared only against Claude Code, not OpenAI's Codex or Google's Gemini CLI. Independent leaderboards show OpenAI's Codex CLI running GPT-5.5 at 82.2 percent on Terminal Bench 2, roughly nine points above MiMo Code's self-reported 73 percent. On SWE-Bench Pro, however, OpenAI reports GPT-5.5 at 58.6 percent, below MiMo Code's claimed 62 percent.

Perhaps more telling: Xiaomi says it ran a double-blind A/B evaluation during internal beta with 576 developers across 474 private repositories. Under 200 execution steps, the two systems split roughly evenly, but past 200 steps MiMo Code's win rate rose above 65 percent—supporting the thesis that memory architecture pays off specifically on long-horizon work.

Aggressive pricing and ecosystem play

MiMo Code ships with zero-configuration access to MiMo-V2.5, the 310-billion-parameter sparse mixture-of-experts model Xiaomi released in April 2026. The model is priced at $0.40 per million input tokens and $2.00 per million output tokens, making it among the cheapest frontier models available. The larger MiMo-V2.5-Pro runs $1.00 to $3.00 per million tokens depending on context length.

The tool also supports third-party backends including DeepSeek, Moonshot's Kimi, and any OpenAI-compatible API, mirroring the bring-your-own-model flexibility of its OpenCode parent.

Enterprise considerations

For engineering leaders, MiMo Code presents a low-barrier evaluation opportunity: MIT licensing permits modification and commercial use, the OpenCode lineage means the architecture is inspectable, and bring-your-own-model support allows routing to internally approved endpoints. The persistent memory system addresses a widely felt pain point in agentic workflows.

The countervailing factors: free model access is explicitly temporary and routes code context through Xiaomi's servers, which will be a non-starter for organizations with strict data-residency policies. The benchmark edge over Claude Code is self-reported and hasn't been independently verified. A V0.1.0 release number signals early-stage maturity. Teams subject to U.S. government procurement restrictions on Chinese technology vendors should weigh that context before adopting.

These details were first reported by VentureBeat.