MiMo-V2.5-Pro

Xiaomi·2026-04-27

LLMscodingcloud

Overview

Xiaomi's flagship coding/agentic open-weights model. 1.02T total / 42B active hybrid MoE (70 layers, 384 routed experts, 8 active per token) with interleaved Sliding Window + Global Attention at a 6:1 ratio and 128-token window for long-context efficiency. 1M context, MIT licensed, day-0 SGLang/vLLM support. Reports SWE-bench Pro 57.2, GDPVal-AA Elo 1581, ClawEval 64% Pass³ at ~70K tokens per trajectory, and a SysY compiler-in-Rust task solved 233/233 in 4.3h with 672 tool calls. Project lead Fuli Luo (ex-DeepSeek).

Capabilities and innovations

1.02T total / 42B active (hybrid MoE)1M context windowHybrid SWA + Global Attention (6:1 ratio, 128-token window)FP8 E4M3 mixed precisionDay-0 SGLang/vLLM supportMIT licenseInterleaved Sliding Window + Global Attention (6:1)Lightweight MTP modules with dense FFNsLong-horizon agent trajectories (~70K tokens / 672 tool calls)

Benchmarks

Benchmark	Score	Source
SWE-bench Pro	57.2%	third party SWE-bench Pro 57.2 reported via VentureBeat (Build Fast with AI as secondary source). Not corroborated on mimo.xiaomi.com or HuggingFace model card at intake time.
MMLU-Pro	84.59%	unsourced

GPQA Diamond: not reported by the vendor
Humanity’s Last Exam (no tools): not reported by the vendor
SWE-bench Verified: not reported by the vendor
MMLU: not reported by the vendor
Terminal-Bench 2.x: not reported by the vendor
MMMU-Pro: not reported by the vendor

API pricing

$1 per million input tokens, $3 per million output tokens (USD, cheapest OpenRouter endpoint, checked 2026-04-29)

Architecture and hardware

Parameters: 1020B total, 42B active per token (MoE)

AI Model Timeline