Model

MiniMax-Text-01

OPEN

MiniMax · MiniMax-01 family · released Jan 15, 2025

456B-param MoE (45.9B active) using Lightning Attention; trained to 1M-token context and extrapolates to 4M at inference; released open-weight alongside vision variant MiniMax-VL-01.

ReasoningCodingVisionFunction callingTool useAgentic

1288.1

Elo · rank #159

Parameters

456B

Active params

45.9B (MoE)

Context

1M tokens

Architecture

Hybrid Lightning Attention + Softmax Attention, MoE (32 experts, top-2 routing, 80 layers)

License

MiniMax Model License Agreement (custom, source-available; code under MIT)

Languages

—

API price (in/out)

No hosted API

Modalities

text

Benchmark results

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

Arena-HardHuman preference89.1%#14

best: Qwen3-235B-A22B (Non-Thinking) · 96.1%

DROPReasoning87.8#11

best: Hunyuan-T1 · 93.1

GPQA DiamondReasoning54.4%#179

best: GPT-5.6 · 94.6%

GSM8KMath94.8%#12

best: Llama 3.1 405B · 96.8%

HumanEvalCoding86.9%#37

best: Claude Opus 4.5 · 99.4%

IFEvalReasoning89.1%#28

best: Gemma 4 26B A4B · 98.5%

MATH-500Math77.4%#72

best: GPT-5 · 99.4%

MBPP+Coding71.7%#13

best: Llama 3.1 405B · 88.6%

MMLU-ProKnowledge75.7%#73

best: Claude Fable 5 · 91.5%

MMLUKnowledge88.5%#23

best: OpenAI o3 · 92.9%

SimpleQAKnowledge23.7%#21

best: GPT-4.5 · 62.5%

Run it locally

VRAM @ Q4

—

VRAM @ FP16

—

Fits on (Q4)

Multi-node cluster required

MiniMax's own report targets ~640GB memory across 8 GPUs at 8-bit quantization for 1M-token inference; far beyond a single RTX 4090's 24GB, so no consumer-GPU throughput figures exist.

Quantizations

INT8 (Quanto)

MiniMax-01 family

Elo progression across releases

MiniMax-VL-01Jan 20251276.1 MiniMax-Text-01Jan 20251288.1

Resources

HFMiniMaxAI/MiniMax-Text-01 GHMiniMax-AI/MiniMax-01 DOCSplatform.minimax.io/docs APIAvailable via MiniMax API

API price weights · each benchmark row carries its own source badge (see methodology)