Model

MiniMax-M2

OPEN

MiniMax · MiniMax-M2 family · released Oct 27, 2025

Compact 229.9B-total/9.8B-active MoE tuned for coding and agentic workflows with interleaved <think> reasoning; original M2 release is MIT-licensed (later M2.x point releases, e.g. M2.1/M2.7, moved to a restrictive 'Modified-MIT').

ReasoningCodingVisionFunction callingTool useAgentic

1871.5

Elo · rank #73

Parameters

229.9B

Active params

9.8B (MoE)

Context

192K tokens

Architecture

62-layer decoder-only Transformer, 256 experts (top-8, sigmoid gating) + Multi-Token Prediction module, full GQA attention

License

MIT

Languages

—

API price (in/out)

$0.3 / $1.2

Modalities

text

Benchmark results

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

AIMEMath78.0%#85

best: GPT-5.2 · 100.0%

BrowseCompAgents44.0%#34

best: Kimi K3 · 91.2%

GAIAAgents75.7%#1

best: this model · 75.7%

GPQA DiamondReasoning78.0%#90

best: GPT-5.6 · 94.6%

Humanity's Last ExamReasoning31.8%#42

best: Claude Sonnet 5 · 57.4%

IFBenchReasoning72.0%#24

best: MiniMax M3 · 83.0%

LiveCodeBenchCoding83.0%#35

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

MMLU-ProKnowledge82.0%#45

best: Claude Fable 5 · 91.5%

SWE-bench VerifiedCoding69.4%#61

best: Claude Fable 5 · 95.0%

τ²-Bench TelecomAgents87.0%#22

best: Claude Opus 4.6 · 99.3%

Terminal-BenchCoding46.3%#3

best: Claude Opus 4.5 (High) · 59.3%

Run it locally

VRAM @ Q4

130 GB

VRAM @ FP16

460 GB

Fits on (Q4)

M3 Ultra 512GBH200 141GBB200 192GB

Unsloth's UD-IQ4_XS dynamic GGUF (~108GB) is reported at ~15-25 tok/s on a single 16GB GPU + 96GB system RAM or Mac unified memory - no RTX-4090-specific figure found.

Quantizations

GGUF · AWQ

MiniMax-M2 family

Elo progression across releases

MiniMax-M2Oct 20251871.5

Resources

HFMiniMaxAI/MiniMax-M2 GHMiniMax-AI/MiniMax-M2 DOCSplatform.minimax.io APIAvailable via MiniMax API

API price $0.3/$1.2 · each benchmark row carries its own source badge (see methodology)