Summary
Xiaomi's MiMo-V2.5-Pro is a trillion-parameter mixture-of-experts model (42B active) with a 1M token context window, positioned as a major advance in agentic intelligence and long-horizon consistency. Demonstrations include building a complete SysY compiler in Rust in 4.3 hours with 672 tool calls, scoring 233/233 on hidden tests; creating an 8,192-line desktop video editor in 11.5 hours; and designing an FVF-LDO analog circuit in about an hour on TSMC 180nm, meeting all six performance targets. The model scores 57.2 on coding agent benchmarks, 73.7 on SWE-Bench Pro, and reportedly uses 40-60% fewer tokens than competitors on comparable trajectories. It supports Claude Code, OpenCode, and Kilo ecosystems, with input pricing of $1 per million tokens and $3 for output. Xiaomi plans to release official weights, and the smaller MiMo-V2.5 multimodal variant costs half as much.
Xiaomi's MiMo-V2.5-Pro (announced April 22, 2026) is presented by the company as a major leap in agentic intelligence and long-horizon consistency — an AI system capable of independently carrying out tasks spanning thousands of tool calls while maintaining logical coherence.
Architecture and Capabilities
- ~1 trillion parameter MoE architecture with 42B active parameters
- 1M token context window with a hybrid attention mechanism for long-range dependencies
- Designed for self-correction, memory management, and goal persistence across extended tasks
Headline Demos
SysY Compiler (Rust)
On Peking University's classic Compiler Principles project — building a full SysY compiler from scratch — the model completed the task in 4.3 hours with 672 tool calls, scoring 233/233 on the hidden test suite. Its first compile passed 137/233 (59% cold-start). It solved Koopa IR (110/110), the RISC-V backend (103/103), and performance optimization (20/20), and diagnosed and fixed a regression at iteration 512.Desktop Video Editor
From a few prompt sentences, it produced an 8,192-line desktop video editor in 11.5 hours with 1,868 tool calls, featuring multi-track timelines, clip trimming, crossfades, audio mixing, AI voiceover (MiMo-V2-TTS), and a full export pipeline.FVF-LDO Analog Circuit Design
On a graduate-level EDA task — designing an FVF-LDO on TSMC 180nm — the model iterated within an ngspice closed loop in about one hour, meeting all six targets (phase margin, line/load regulation, quiescent current, PSRR, transient response), with some metrics improving by an order of magnitude.Benchmark Results
- Coding Agent: 57.2 (vs Claude Opus 4.6 at 57.3, GPT-5.4 at 57.7, previous MiMo-V2-Pro at 55.0)
- SWE-Bench Pro: 73.7
- MiMo Coding Bench: 68.4 (previous generation: 57.1)
- Terminal-Bench 2.0: third place; FrontierSWE: 1581; GDPVal-AA: 72.9; τ3-bench: 63.8
- Claw-Eval (Pass@3): 48.0 without tools, 34.0 with tools
- Xiaomi claims 40–60% token savings vs competitors on comparable trajectories
Availability
- Pro pricing: $1/million input tokens, $3/million output tokens
- MiMo-V2.5 (multimodal, non-Pro) priced at half of Pro
- Compatible with Claude Code, OpenCode, and Kilo agent ecosystems
- Xiaomi plans an official release with open-sourced 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
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/177618671