English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Xiaomi MiMo-V2.5-Pro: A Leap in Agentic and Long-Horizon Coherence

Forum topic · ✨步子哥 · 2026-04-23

Summary

Xiaomi introduced MiMo-V2.5-Pro on April 22, 2026, describing it as a leap forward in agentic capability and long-horizon coherence. The model sustains chains of over 1,000 tool calls while preserving logic and instruction fidelity. Official demos show it building a SysY compiler in Rust from scratch in 4.3 hours (672 tool calls, eventually passing all 233 hidden tests), creating a full desktop video editor in 11.5 hours with 8,192 lines of code and 1,868 tool calls including MiMo-V2-TTS-powered AI narration, and optimizing an FVF-LDO analog circuit on TSMC 180nm CMOS in about one hour with ngspice closed-loop iteration, improving four of six specs by an order of magnitude. Benchmarks place it near frontier models—around 57.2 on SWE-Bench-related coding-agent tests, 68.4 on MiMo Coding Bench, 72.9 on GDPVal-AA—while using 40-60% fewer tokens than competitors. The hybrid-attention model supports long context, runs at 60-80 tokens/s, is priced at $1/M input and $3/M output tokens, and its weights are planned to be open-sourced.

Xiaomi launched MiMo-V2.5-Pro on April 22, 2026, officially describing it as "a leap in agentic and long horizon coherence." Rather than chasing leaderboard scores alone, the model targets real-world tasks requiring hundreds or thousands of tool calls over hours or days.

Long-Horizon Coherence

While traditional large models tend to lose coherence as context grows, MiMo-V2.5-Pro sustains task chains of over 1,000 tool calls while maintaining logical rigor and instruction fidelity. It is particularly adept at following subtle contextual instructions hidden within agentic scenarios.

Official Demos

SysY Compiler in Rust — 4.3 hours

Based on a Peking University compiler course project, the model independently completed lexical analysis, parsing, AST construction, Koopa IR generation, a RISC-V assembly backend, and optimization:

  • 4.3 hours, 672 tool calls
  • First full compile passed 137/233 hidden tests (59% cold-start pass rate)
  • Final result: 233/233
  • Koopa IR: 110/110; RISC-V backend: 103/103; optimization: 20/20
  • The model demonstrated structured self-correction discipline: it planned milestones and validated intermediate results instead of blind trial-and-error.

    Desktop Video Editor — 11.5 hours

    Given instructions to implement a multi-track timeline, clip cutting, cross-fades, audio mixing, and an export pipeline:

  • 11.5 hours, 8,192 lines of code, 1,868 tool calls
  • Integrated AI voice narration powered by MiMo-V2-TTS
  • Showed strong "harness awareness," collaborating seamlessly with external frameworks
  • Analog Circuit Optimization — ~1 hour

    Designing and optimizing an FVF-LDO (flipped voltage follower low-dropout regulator) under TSMC 180nm CMOS, meeting six specifications (phase margin, line/load regulation, quiescent current, PSRR, transient response):

  • Closed-loop iteration with ngspice simulation
  • All six specs met in about 1 hour
  • Four specs improved by an order of magnitude over the initial design
  • Benchmarks and Efficiency

    Key reported scores include:

  • Coding agent / SWE-Bench related: 57.2 (near Claude Opus 4.6's 57.3 and GPT-5.4's 57.7); up to 73.7 on a SWE-Bench Pro variant
  • MiMo Coding Bench: 68.4 (vs. Claude Opus 4.6's 65.4)
  • GDPVal-AA: 72.9 (Elo 1581)
  • τ3-bench: 63.8
  • Claw-Eval (Pass@3): 34.0 with tools / 48.0 without; ~64% pass rate at roughly 70K tokens per trajectory
  • The model uses 40-60% fewer tokens to achieve comparable results, positioning it in the ideal high-score/low-cost region and making it attractive for integration into agent frameworks like Claude Code, OpenCode, and Kilo.

    Architecture and Access

  • Hybrid attention mechanism with very long context (up to 1M tokens in the series)
  • Speed of roughly 60-80 tokens/s
  • API pricing: $1/M input tokens, $3/M output tokens
  • Public beta available; multimodal capabilities (image, audio, video) in the MiMo-V2.5 series
  • Open-sourced weights planned at official release

Conclusion

MiMo-V2.5-Pro demonstrates that higher intelligence means completing harder tasks with fewer resources—moving AI from a chat tool toward a genuine collaborator capable of building software, media tools, and hardware designs end-to-end.

Tags

#xiaomi#mimo-v25-pro#ai-agents#long-horizon-tasks#tool-calling#benchmarks#llm#open-source

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177618670