Model

MiniMax-M1

OPEN

MiniMax · MiniMax-M1 family · released Jun 16, 2025

World's first open-weight large-scale hybrid-attention reasoning model; shipped as two checkpoints (40K/80K thinking budget) - this entry reflects the flagship 80K checkpoint.

ReasoningCodingVisionFunction callingTool useAgentic

1547.8

Elo · rank #129

Parameters

456B

Active params

45.9B (MoE)

Context

1M tokens

Architecture

Hybrid Lightning+Softmax Attention MoE (80 layers, 32 experts top-2) reasoning model, trained with CISPO RL

License

Apache-2.0

Languages

—

API price (in/out)

$0.4 / $2.2

Modalities

text

Benchmark results

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

AIMEMath86.0%#65

best: GPT-5.2 · 100.0%

GPQA DiamondReasoning69.2%#134

best: GPT-6 Astra · 96.0%

Humanity's Last ExamReasoning8.4%#102

best: Claude Opus 5 · 64.7%

LiveCodeBenchCoding65.0%#86

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

MATH-500Math96.8%#21

best: GPT-5 · 99.4%

MMLU-ProKnowledge80.6%#54

best: Claude Fable 5 · 91.5%

SimpleQAKnowledge18.5%#25

best: GPT-4.5 · 62.5%

SWE-bench VerifiedCoding56.0%#88

best: Claude Opus 5 · 96.0%

tau-benchAgents63.5%#11

best: Claude Opus 4.1 · 82.4%

Run it locally

VRAM @ Q4

—

VRAM @ FP16

912 GB

Fits on (Q4)

Multi-node cluster required

llama.cpp did not support the MiniMaxM1ForCausalLM architecture at release, so no GGUF/Q4 quant or RTX 4090 figure exists.

Quantizations

—

Fine-tune it

Permissive

QLoRA310.1 GB2× B200 192GB

LoRA971.3 GB8× H200 141GB

Full fine-tune7318.8 GBbeyond 8× B200

QLoRA SFT on ~10k samples ≈ $163 (2× B200 192GB)

Plan a fine-tune →

MiniMax-M1 family

Elo progression across releases

MiniMax-M1Jun 20251547.8

Resources

HFMiniMaxAI/MiniMax-M1-80k GHMiniMax-AI/MiniMax-M1 DOCSplatform.minimax.io APIAvailable via MiniMax API

API price $0.4/$2.2 · each benchmark row carries its own source badge (see methodology)