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

Xiaomi released an open-source terminal-native coding agent called MiMo Code on June 10, 2026, under an MIT license. The company claims MiMo Code outperforms Anthropic's Claude Code on long-horizon workflows of 200+ steps, with a reported win rate above 65% at 200 execution steps. The agent pairs with the multimodal model MiMo-V2.5 and features a four-layer cross-session memory architecture.

Xiaomi MiMo Code claims edge over Claude Code on 200-step tasks VentureBeat and Xiaomi's MiMo blog report that Xiaomi published an open-source, terminal-native coding agent called MiMo Code on June 10, 2026. VentureBeat reports the announcement included an internal beta and a survey of 576 developers , and Xiaomi published the project under an MIT license on GitHub. Multiple outlets Indian Express, Developer-Tech report Xiaomi's benchmark results showing MiMo Code outperforming Anthropic's Claude Code on long-horizon, multi-step workflows of 200+ steps , with one report citing a win rate above 65% at 200 execution steps. VentureBeat and Decrypt further report MiMo Code pairs the agent harness with the multimodal model MiMo-V2.5, which the company offered as limited-time free access and which Decrypt describes as much faster than ChatGPT and Claude. VentureBeat documents MiMo Code's four-layer cross-session memory built on SQLite FTS5. What happened VentureBeat reports Xiaomi's MiMo team released an open-source, terminal-native coding agent called MiMo Code on June 10, 2026. Per VentureBeat and Xiaomi's blog post, the project is distributed under an MIT license on GitHub and installs via a single terminal command or npm for Windows. VentureBeat reports the announcement referenced an internal beta plus a survey of 576 developers . Multiple outlets Indian Express, Developer-Tech, GadgetsNow report Xiaomi's benchmark comparisons against Anthropic's Claude Code , showing MiMo Code completing long-horizon, multi-step developer workflows of 200+ steps with a reported win rate above 65% at 200 execution steps in Xiaomi's published results. VentureBeat and Decrypt report Xiaomi paired the agent harness with the multimodal model MiMo-V2.5, which Decrypt characterises as significantly faster than ChatGPT and Claude, and Xiaomi offered limited-time free access to that model. Technical details reported VentureBeat reports MiMo Code is a fork of the open-source OpenCode agent extended with a memory architecture, workflow modes, and a model harness. VentureBeat documents a four-layer cross-session memory implemented with SQLite FTS5 full-text search; the layers are listed in the announcement as project memory a persistent MEMORY.md , session checkpoints, scratch notes, and per-task progress logs. VentureBeat reproduces a direct quote from the MiMo post describing the tool as "more than an AI coding assistant in your terminal, it's the smartest coding partner you'll ever work with." Editorial analysis - technical context Observed patterns in similar agentic systems show that long-horizon workflows stress both context-window limits and state-management. Industry-pattern observations: agent performance on multi-step developer tasks typically degrades when earlier task state is compressed or discarded; persistent memory plus retrieval is a common mitigation strategy. For practitioners: a four-layer memory stack combined with FTS5 search suggests Xiaomi prioritised inexpensive, local persistence and text-indexed retrieval rather than depending solely on expanded context windows or model-side compression. Context and significance public comparisons between vendor-provided agents Xiaomi and incumbents Anthropic's Claude Code , OpenAI-based assistants are increasingly common as vendors attempt to demonstrate robustness on longer workflows. Observed patterns in benchmark reporting include reliance on in-house test suites and curated task sets; when claimed wins come from vendor-run evaluations, independent verification is required to assess generalisability. For developers and tooling teams, the combination of an open-source agent harness plus a documented memory architecture lowers the barrier to reproduce or extend the approach in private projects. What to watch - •Independent evaluations replicating Xiaomi's 200-step tasks and reporting reproducible win rates against Claude Code and other agents. - •Community uptake and forks on GitHub that test MiMo-V2.5 performance across languages and real-world repos. - •Resource and latency trade-offs when combining MiMo-V2.5 million-token context claims in Xiaomi's posts reported by VentureBeat with a persistent SQLite FTS5 store in CI and developer machines. All reported performance figures, the 576-developer survey, the 65% + win-rate at 200 steps , and the MiMo-V2.5 free-access claim are drawn from Xiaomi's announcement as covered by VentureBeat, Indian Express, Developer-Tech, GadgetsNow, and Decrypt. Xiaomi has not provided, in the cited sources, an independent third-party benchmark or public leaderboard results beyond the published materials cited above. Scoring Rationale Notable product release: an open-source agent plus a million-token-capable model and a documented cross-session memory are useful to practitioners. The core claims are vendor-run benchmarks needing independent verification, which reduces immediate industry-shaking impact. Practice interview problems based on real data 1,500+ SQL & Python problems across 15 industry datasets — the exact type of data you work with. Try 250 free problems /problems