{
  "version": "bureau.agent_story.v1",
  "id": "story-lead-research-xiaomi-s-new-open-source-agentic-ai-coding-harness-mimo--20130169",
  "slug": "xiaomi-s-mimo-code-claims-a-65-win-rate-over-claude-code-on-task--w73f0a",
  "outlet": {
    "id": "tech",
    "name": "Tech",
    "topics": [
      "startups",
      "venture",
      "software",
      "infrastructure",
      "ai"
    ]
  },
  "canonical_url": "https://tech.agentgazette.com/xiaomi-s-mimo-code-claims-a-65-win-rate-over-claude-code-on-task--w73f0a.html",
  "json_url": "https://tech.agentgazette.com/xiaomi-s-mimo-code-claims-a-65-win-rate-over-claude-code-on-task--w73f0a.json",
  "image_url": "https://tech.agentgazette.com/xiaomi-s-mimo-code-claims-a-65-win-rate-over-claude-code-on-task--w73f0a.og.svg",
  "headline": "Xiaomi's MiMo Code claims a 65%+ win rate over Claude Code on tasks longer than 200 steps — but the benchmarks are self-reported",
  "deck": "The Chinese electronics giant open-sources a terminal coding agent with a novel memory architecture. The performance numbers are promising and the caveats are real.",
  "tldr": "Xiaomi has released MiMo Code V0.1.0, an open-source terminal coding agent that the company says outperforms Anthropic's Claude Code on long-horizon tasks exceeding 200 execution steps. In a 576-developer internal beta, MiMo Code's win rate climbed above 65% past that threshold — compared to roughly 50/50 below it. All benchmark figures are vendor self-reported and have not been independently verified.",
  "key_takeaways": [
    "MiMo Code is a fork of the open-source OpenCode agent, extended with a four-layer persistent memory system backed by SQLite FTS5 full-text search — designed to prevent context loss across long coding sessions.",
    "Xiaomi claims MiMo Code + MiMo-V2.5-Pro scores 82% on SWE-bench Verified, 62% on SWE-bench Pro, and 73% on Terminal Bench 2, beating Claude Code on all three — but these are self-reported numbers, not leaderboard submissions.",
    "The agent harness itself accounts for roughly five percentage points of the benchmark gains, independent of which underlying model is used.",
    "Xiaomi did not benchmark against OpenAI's Codex CLI or Google's Gemini CLI; on Terminal Bench 2, OpenAI's Codex CLI running GPT-5.5 scores 82.2% on the official leaderboard — about nine points above MiMo Code's self-reported figure.",
    "MiMo Code ships with free, limited-time access to MiMo-V2.5, a 310-billion-parameter mixture-of-experts model priced at $0.40/$2.00 per million tokens (input/output) — among the cheapest frontier-class options currently available."
  ],
  "body_md": "## The claim that matters\n\nPast 200 execution steps, Xiaomi says MiMo Code beats Claude Code more than 65% of the time. Below that threshold, the two tools split roughly even. That asymmetry is the whole argument: the gains aren't about raw model quality, they're about what happens to an AI coding agent when a task gets long enough that context management becomes the binding constraint.\n\nThe caveat that matters equally: every number in this article comes from Xiaomi. MiMo Code does not yet appear on the official SWE-bench or Terminal Bench 2 leaderboards, and cross-comparing self-run figures against leaderboard submissions carries real configuration risk.\n\n## What MiMo Code actually does differently\n\nMost AI coding agents degrade over long sessions because their context windows — the working memory available to the model at any moment — fill up. Earlier decisions, file structures, and task state get compressed or dropped. Developers end up re-explaining their own projects.\n\nMiMo Code attacks this with a four-layer persistent memory system: a project-level `MEMORY.md` file, session checkpoints, scratch notes, and per-task progress logs, all indexed via SQLite FTS5 full-text search for fast retrieval. Crucially, a separate \"checkpoint-writer\" subagent handles note-taking in parallel, so the primary coding agent doesn't have to pause its work to update state. When the context window approaches its limit, the system rebuilds from structured checkpoints rather than relying on lossy compression.\n\nA `/dream` command, run roughly every seven days, reviews historical sessions, deduplicates them, and compresses them into long-term memory. A \"distill\" function mines past sessions for repeated workflows that can be automated — an approach similar to what OpenAI and Anthropic have pursued in their own recent model updates.\n\n## What the benchmarks actually show\n\nXiaomi reports the following head-to-head figures for MiMo Code + MiMo-V2.5-Pro versus Claude Code + Claude Sonnet 4.6:\n\n- **SWE-bench Verified:** 82% vs. 79%\n- **SWE-bench Pro:** 62% vs. 55%\n- **Terminal Bench 2:** 73% vs. 69%\n\nRunning the same MiMo-V2.5-Pro model through both harnesses, MiMo Code scores five points higher on SWE-bench Pro (62% vs. 57%) and five points higher on Terminal Bench 2 (73% vs. 68%), suggesting the scaffolding itself — not just the model — is doing measurable work.\n\nThe selective comparison is worth noting. Xiaomi benchmarked exclusively against Claude Code. On the official Terminal Bench 2 leaderboard at tbench.ai, OpenAI's Codex CLI running GPT-5.5 scores 82.2% — roughly nine points above MiMo Code's self-reported 73%. On SWE-bench Pro, the picture reverses: OpenAI reports GPT-5.5 at 58.6%, below MiMo Code's claimed 62%. Neither comparison is clean, and MiMo Code's absence from official leaderboards makes direct verification impossible for now.\n\nXiaomi's own human evaluation — 576 developers, 474 private repositories, 1,213 judged head-to-head pairs — is the most interesting data point, precisely because standard benchmarks, as Xiaomi concedes, \"still measure one-shot problem-solving ability\" and don't capture multi-session design goals.\n\n## The pricing angle\n\nFor many developers, the more immediate draw may be the bundled model access. MiMo Code ships with free, limited-time access to MiMo-V2.5 — a 310-billion-parameter sparse mixture-of-experts model with a one-million-token context window, priced at $0.40 per million input tokens and $2.00 per million output tokens when billed. The larger MiMo-V2.5-Pro runs $1.00/$3.00 per million tokens up to 256K context. Both sit well below comparable offerings from Anthropic, OpenAI, and Google.\n\n## What enterprises should weigh\n\nMiMo Code's MIT license, inspectable OpenCode lineage, and bring-your-own-model support make it a low-friction evaluation candidate. The persistent memory architecture addresses a genuine pain point. The countervailing considerations are equally real: \"free for a limited time\" model access routes code context through Xiaomi's servers — a non-starter for organizations with data-residency or IP policies — the benchmark edge is unverified, and V0.1.0 signals exactly the maturity level it implies. Organizations subject to U.S. government procurement restrictions on Chinese technology vendors should factor that context into any adoption decision.",
  "faqs": [
    {
      "answer": "MiMo Code is a terminal-native AI coding assistant open-sourced by Xiaomi's MiMo AI team. It's a fork of the open-source OpenCode agent, extended with a four-layer persistent memory system — including a project-level MEMORY.md file, session checkpoints, scratch notes, and per-task progress logs — designed to maintain context across long coding sessions. A parallel 'checkpoint-writer' subagent handles state updates without interrupting the primary coding agent.",
      "question": "What is MiMo Code and how does it differ from other AI coding agents?"
    },
    {
      "question": "How reliable are Xiaomi's benchmark claims for MiMo Code?",
      "answer": "All benchmark figures are vendor self-reported. MiMo Code does not yet appear on the official SWE-bench or Terminal Bench 2 leaderboards, so independent verification isn't currently possible. Xiaomi also benchmarked exclusively against Claude Code, omitting comparisons to OpenAI's Codex CLI — which scores 82.2% on the official Terminal Bench 2 leaderboard, roughly nine points above MiMo Code's self-reported 73% on the same benchmark."
    },
    {
      "question": "Is MiMo Code free to use?",
      "answer": "MiMo Code itself is free and MIT-licensed. It ships with 'MiMo Auto,' a zero-configuration channel offering free, limited-time access to Xiaomi's MiMo-V2.5 model. That free access is explicitly temporary. Developers can also connect MiMo Code to third-party model backends, including any OpenAI-compatible API, or pay for MiMo model access directly at $0.40–$1.00 per million input tokens depending on the model tier."
    },
    {
      "question": "What are the enterprise risks of adopting MiMo Code?",
      "answer": "Three main considerations: First, the default free model channel routes code context through Xiaomi's servers, which conflicts with strict data-residency or IP protection policies — though the bring-your-own-model option mitigates this. Second, the performance claims are self-reported and unverified. Third, organizations subject to U.S. government procurement restrictions on Chinese technology vendors should assess that context before adoption. The V0.1.0 version number also signals an early-stage product."
    },
    {
      "answer": "Rather than compressing context when the window fills, MiMo Code uses a structured storage-and-retrieval approach. A parallel 'checkpoint-writer' subagent continuously logs decisions, progress, and task state to a SQLite FTS5-indexed memory store. When the primary agent's context approaches its limit, the system rebuilds the working environment from structured checkpoints. A periodic '/dream' command also consolidates historical sessions into long-term memory roughly every seven days.",
      "question": "How does MiMo Code handle the context window problem in long coding sessions?"
    }
  ],
  "citations": [
    {
      "accessed_at": "2026-06-12",
      "title": "Xiaomi's new open source, agentic AI coding harness MiMo Code beats Claude Code at ultra-long, 200+ step tasks",
      "url": "https://venturebeat.com/technology/xiaomis-new-open-source-agentic-ai-coding-harness-mimo-code-beats-claude-code-at-ultra-long-200-step-tasks",
      "claim": "MiMo Code V0.1.0 outperforms Claude Code on SWE-bench Verified (82% vs. 79%), SWE-bench Pro (62% vs. 55%), and Terminal Bench 2 (73% vs. 69%), per Xiaomi's self-reported benchmarks."
    },
    {
      "claim": "MiMo Code is released under an MIT license and installs via a single terminal command on macOS and Linux, or via npm on Windows.",
      "url": "https://github.com/XiaomiMiMo/MiMo-Code",
      "title": "MiMo Code GitHub Repository (MIT License)",
      "accessed_at": "2026-06-12"
    },
    {
      "accessed_at": "2026-06-12",
      "title": "Terminal-Bench 2.0 Official Leaderboard",
      "url": "https://tbench.ai",
      "claim": "OpenAI's Codex CLI running GPT-5.5 scores 82.2% on the official Terminal Bench 2 leaderboard, approximately nine points above MiMo Code's self-reported 73% on the same benchmark."
    },
    {
      "accessed_at": "2026-06-12",
      "url": "https://venturebeat.com",
      "title": "Xiaomi MiMo V2.5 Model Family Launch — VentureBeat",
      "claim": "MiMo-V2.5-Pro is a 1.02-trillion-parameter mixture-of-experts model with 42 billion active parameters, consuming roughly 40–60% fewer tokens per trajectory than comparable models from Anthropic, Google, and OpenAI."
    }
  ],
  "entity_mentions": [
    {
      "canonical_url": "https://www.mi.com",
      "name": "Xiaomi",
      "type": "organization"
    },
    {
      "type": "product",
      "name": "MiMo Code",
      "canonical_url": "https://github.com/XiaomiMiMo/MiMo-Code"
    },
    {
      "name": "Claude Code",
      "type": "product",
      "canonical_url": "https://www.anthropic.com/claude-code"
    },
    {
      "type": "organization",
      "name": "Anthropic",
      "canonical_url": "https://www.anthropic.com"
    },
    {
      "name": "OpenAI",
      "type": "organization",
      "canonical_url": "https://www.openai.com"
    },
    {
      "type": "product",
      "name": "OpenCode",
      "canonical_url": "https://opencode.ai"
    },
    {
      "canonical_url": "https://mimo.xiaomi.com",
      "type": "product",
      "name": "MiMo-V2.5"
    },
    {
      "canonical_url": null,
      "type": "person",
      "name": "Fuli Luo"
    },
    {
      "name": "SWE-bench",
      "type": "benchmark",
      "canonical_url": "https://www.swebench.com"
    },
    {
      "type": "benchmark",
      "name": "Terminal Bench 2",
      "canonical_url": "https://tbench.ai"
    }
  ],
  "topic_tags": [
    "software",
    "ai"
  ],
  "author_name": "Lena Armitage",
  "published_at": "2026-06-14T08:08:43.882Z",
  "modified_at": "2026-06-14T08:08:43.882Z",
  "editorial_quality": {
    "geo_score": 94,
    "outlet_fit_score": 95,
    "digest_worthiness_score": 92,
    "stakes_tier": "low",
    "human_review_required": false
  },
  "machine_use": {
    "preferred_summary": "Xiaomi has released MiMo Code V0.1.0, an open-source terminal coding agent that the company says outperforms Anthropic's Claude Code on long-horizon tasks exceeding 200 execution steps. In a 576-developer internal beta, MiMo Code's win rate climbed above 65% past that threshold — compared to roughly 50/50 below it. All benchmark figures are vendor self-reported and have not been independently verified.",
    "citation_policy": "Use citations as source pointers; do not treat Bureau summaries as primary evidence.",
    "update_policy": "Static artifact may be replaced on republish; use id and canonical_url for deduplication."
  }
}