Xiaomi's MiMo Code Benchmarks Past Claude Code on Terminal Bench — as Open Source

Xiaomi released MiMo Code on June 10, 2026 — an open-source agentic coding agent built on MiMo-V2.5-Pro, a 1.02 trillion-parameter mixture-of-experts model with a 1 million token context window. The benchmark headline is a Terminal-Bench 2.0 score of 86.7%, compared to Claude Code’s 65.4% on the same benchmark. That gap, if it holds under independent verification, is the story.

The Number

The 86.7% vs 65.4% comparison is on Terminal-Bench 2.0, a benchmark designed for multi-step terminal-based coding tasks. MiMo Code also posts 57.2% on SWE-bench Pro. Alongside the benchmark data, Xiaomi ran a human A/B test with 576 developers focused specifically on long-horizon performance — tasks running past 200 steps. MiMo Code achieved a 65%+ win rate in those evaluations.

What It Actually Does Differently

The long-horizon performance claim is the interesting claim to examine. Most coding agents degrade sharply past 50–100 steps as context windows fill and compounding errors accumulate. MiMo Code’s architecture directly targets this: it uses a 4-layer cross-session memory system built on SQLite with full-text search (FTS5), designed to maintain coherent state across extended runs.

The practical result is 40–60% fewer tokens consumed at long task horizons compared to agents that reconstruct context from scratch. That’s both a cost advantage and a capability advantage — agents that burn tokens faster hit context limits sooner, compressing the effective task horizon even on models with nominally large windows.

The Open Source Angle

MiMo Code ships as open source — the model weights and agent architecture are publicly available. That’s meaningful for teams evaluating agentic coding infrastructure: an open-weights 1T-parameter model at these benchmark numbers changes the build-vs-buy calculus for organizations that would otherwise be locked into hosted API pricing for long-horizon coding agents.

The SWE-bench Pro score of 57.2% is competitive but not at the top of the leaderboard. The Terminal-Bench 2.0 figure is the stronger claim — and the one that has attracted the most attention. Terminal-Bench 2.0 was specifically designed for multi-step, real-environment coding tasks rather than isolated code generation, which makes it a better proxy for agentic coding performance than single-shot benchmarks.

The Caveat

These are self-reported benchmarks from Xiaomi. The Terminal-Bench 2.0 numbers have not been independently replicated at time of writing. The developer A/B test was conducted by Xiaomi’s team with selected participants. Neither disqualifies the results, but both mean the 86.7% figure should be treated as directional until third parties publish reproducible evaluations on the same benchmark version and configuration.

The model architecture and training details in the technical report appear credible — a 1T-parameter MoE model with 1M context and a purpose-built long-horizon memory system is a coherent explanation for the claimed performance profile. The benchmark numbers are plausible given the architecture. That is different from confirmed.

What It Signals

MiMo Code is one of several signals in 2026 that the frontier for agentic coding is moving from raw model capability to long-horizon reliability. The 200-step threshold Xiaomi uses as a benchmark cutoff reflects where current agents fail in practice — not on simple tasks, but on sustained, multi-file, multi-session engineering work.

The open-source release also continues a pattern: Chinese AI labs releasing capable open-weight models in domains where closed API providers currently charge for access. Whether the benchmarks hold under independent testing determines whether MiMo Code becomes a serious alternative to Claude Code and similar hosted agents, or a benchmark-optimized release that doesn’t transfer to real-world deployment.

Source: VentureBeat

Content created with AI assistance and reviewed for accuracy.