Model

MiMo-V2.5

OPEN

Xiaomi · MiMo family · released Apr 22, 2026

Xiaomi's native omnimodal flagship-tier model (text/image/video/audio), 310B total / 15B active MoE, 1M context, fully open-sourced under MIT. Consolidates the prior separate MiMo-V2-Pro (reasoning) and MiMo-V2-Omni (multimodal) lines into one architecture built on the MiMo-V2-Flash backbone.

ReasoningCodingVisionFunction callingTool useAgentic

2050.2

Elo · rank #75

Parameters

310B

Active params

15B (MoE)

Context

1M tokens

Architecture

Sparse MoE (256 routed experts, 8/token), hybrid Sliding-Window+Global attention 5:1, native 729M-param ViT + 261M-param audio encoder, 3x MTP modules

License

MIT

Languages

—

API price (in/out)

$0.14 / $0.28

Modalities

text · vision · audio · video

Benchmark results

Bar shows position within the tracked field; marker = field best

Agents' Last ExamAgents23.6%#21

best: GPT-6 Astra · 59.3%

Arena EloHuman preference1434#12

best: Claude Fable 5 · 1505

CharXivVision81.0%#14

best: Qwen3.8-Flash-Next · 90.6%

GDPval-AAAgents1145#38

best: Claude Fable 5 · 1932

GPQA DiamondReasoning84.9%#62

best: GPT-6 Astra · 96.0%

Humanity's Last ExamReasoning25.2%#61

best: Claude Opus 5 · 64.7%

IFBenchReasoning67.1%#33

best: MiniMax M3 · 83.0%

LiveCodeBenchCoding81.5%#43

best: DeepSeek-V4-Pro (Think Max) · 93.5%

MMLU-ProKnowledge82.9%#40

best: Claude Fable 5 · 91.5%

MMMU-ProVision77.9%#21

best: Claude Opus 4.7 · 85.5%

PinchBenchAgents89.7%#3

best: Trinity-Large-Thinking · 91.9%

SWE-bench ProCoding56.1%#31

best: Claude Fable 5.1 · 81.2%

SWE-bench VerifiedCoding71.0%#60

best: Claude Opus 5 · 96.0%

τ²-Bench TelecomAgents90.6%#19

best: Claude Opus 4.6 · 99.3%

τ³-BankingAgents6.6%#8

best: GPT-5.6 Sol · 33.0%

Terminal-Bench 2.0Coding65.8%#36

best: Gemini 3.8 Flash · 89.4%

Video-MMEVision87.7%#4

best: Seed 2.1 Pro · 89.2%

Run it locally

VRAM @ Q4

—

VRAM @ FP16

—

Fits on (Q4)

Multi-node cluster required

Throughput data unavailable.

Quantizations

—

Fine-tune it

Permissive

QLoRA210.8 GB2× H200 141GB

LoRA660.3 GB4× B200 192GB

Full fine-tune4975.5 GBbeyond 8× B200

QLoRA SFT on ~10k samples ≈ $7.76 (2× H200 141GB)

Plan a fine-tune →

MiMo family

Elo progression across releases

MiMo-V2.5Apr 20262050.2 MiMo-V2.5-ProApr 20262187.4

Resources

HFXiaomiMiMo/MiMo-V2.5 DOCShttps://mimo.xiaomi.com/mimo-v2-5/APIAvailable via Xiaomi API

API price $0.14/$0.28 · each benchmark row carries its own source badge (see methodology)