Xiaomi MiMo-V2.5-Pro: The Architecture Behind Ten-Thousand-Step Agentic Intelligence

Forum topic · ✨步子哥 · 2026-04-23

Summary

Xiaomi's MiMo-V2.5-Pro, announced April 22, 2026, is positioned by the company as a major leap in agentic intelligence and long-horizon consistency. The model is built on a roughly 1-trillion-parameter Mixture-of-Experts architecture with 42B active parameters, a 1M token context window, and hybrid attention, enabling sustained multi-thousand tool-call workflows. Xiaomi showcases three demos: building a complete SysY compiler in Rust from scratch in 4.3 hours with 672 tool calls, scoring 233/233 on a hidden test suite; creating a full desktop video editor with 8,192 lines of code in 11.5 hours across 1,868 tool calls; and optimizing a TSMC 180nm FVF-LDO analog circuit in about one hour within an ngspice closed loop. Benchmarks include 57.2 on a Coding Agent benchmark (near Claude Opus 4.6's 57.3 and GPT-5.4's 57.7), 73.7 on SWE-Bench Pro, and 68.4 on the MiMo Coding Bench. Xiaomi claims 40-60% token savings versus competitors on similar trajectories. Pricing is $1 per million input tokens and $3 per million output tokens, with a multimodal MiMo-V2.5 at half that price. Open-source weights are planned.

The following is an English translation of a Chinese forum post discussing Xiaomi's MiMo-V2.5-Pro release. The post presents Xiaomi's official claims and demos; figures below are as reported in the source.

Overview

Xiaomi announced MiMo-V2.5-Pro on April 22, 2026, describing it as a "major leap in agentic intelligence and long-horizon consistency." Unlike a chatbot, the model is designed to carry out long task chains — including thousands of tool calls — while maintaining logical coherence and self-correction.

Key architectural claims:

~1 trillion parameters (MoE) with 42B active parameters
1M token context window
Hybrid attention mechanism for long-range dependencies

Demo 1: SysY Compiler in Rust (4.3 hours)

Based on Peking University's classic Compiler Principles project, the model built a complete SysY compiler from scratch — lexer, parser, AST, Koopa IR generation, RISC-V backend, and optimization passes — in 4.3 hours with 672 tool calls, scoring a perfect 233/233 on a hidden test suite.

First-pass compile already passed 137/233 tests (59% cold-start success)
Progressively completed Koopa IR (110/110), RISC-V backend (103/103), and optimizations (20/20)
After a regression at iteration 512, it diagnosed and fixed the issue autonomously — the post highlights this "long-horizon self-correction" as the model's defining trait

Demo 2: Full Video Editor (11.5 hours)

From a few plain-language prompts, MiMo-V2.5-Pro produced 8,192 lines of code in 11.5 hours across 1,868 tool calls. Features included multi-track timeline editing, clip trimming, crossfades, audio mixing, AI voiceover (MiMo-V2-TTS), and a full export pipeline.

Demo 3: FVF-LDO Analog Circuit Optimization (~1 hour)

For a graduate-level analog EDA task — designing and optimizing an FVF-LDO in TSMC 180nm — the model iterated within an ngspice closed-loop simulation for about one hour. It tuned power transistors, compensation networks, and bias voltages to simultaneously satisfy six metrics: phase margin, line regulation, load regulation, quiescent current, PSRR, and transient response. All targets were met, with some metrics improving an order of magnitude over initial attempts.

Benchmarks (as reported)

| Benchmark | Score | |---|---| | Coding Agent | 57.2 (vs Claude Opus 4.6: 57.3, GPT-5.4: 57.7, MiMo-V2-Pro: 55.0) | | SWE-Bench Pro | 73.7 | | MiMo Coding Bench | 68.4 (predecessor: 57.1) | | Terminal-Bench 2.0 | 3rd place | | FrontierSWE | 1581 | | GDPVal-AA | 72.9 | | τ3-bench | 63.8 | | Claw-Eval (Pass@3, no tools / with tools) | 48.0 / 34.0 |

Xiaomi additionally claims the model uses 40-60% fewer tokens than competitors on similar trajectories.

Pricing and Ecosystem

MiMo-V2.5-Pro: $1 per million input tokens, $3 per million output tokens
MiMo-V2.5 (non-Pro multimodal version, vision + audio): half the Pro price
Compatible with agentic frameworks such as Claude Code, OpenCode, and Kilo
Xiaomi plans an official release and open-source weights, enabling self-hosting and fine-tuning

References

1. Xiaomi Official Product Page: MiMo-V2.5-Pro, April 22, 2026, https://mimo.xiaomi.com/mimo-v2-5-pro 2. MiMo-V2.5-Pro Technical Benchmarks and Demo Reports, Xiaomi AI Lab, 2026 3. SysY Compiler Project Reference, Peking University Compiler Principles Course, GitHub pku-minic 4. FVF-LDO Analog Circuit Design Methodology, TSMC 180nm Process Documentation, 2026 Internal Simulation 5. Claw-Eval and SWE-Bench Pro Evaluation Framework, Frontier AI Research Consortium, 2026

*Note: All performance figures and dates reflect the source post's claims and have not been independently verified.*