MiniMax Text 01

MiniMaxAI · released 2025-01-15 · MiniMax Model License license

MiniMax-Text-01 is MiniMax's Jan 2025 456B-total/45.9B-active hybrid-attention MoE built for very long context (up to 4M tokens), with strong long-context retrieval and competitive general benchmarks (MMLU-Pro 75.7, IFEval 89.1).

Key specs

TypeLocal open-weight
Parameters456.09B total · MoE, — active
Architectureminimax_text_01
Context window10240K tokens
Knowledge cutoff2024-03-31
Modalitiestext, vision
Recommended backends
Minimum viable rig

Benchmark scores

GPQA Diamond54.4%
SWE-bench Verified
AIME
MMLU-Pro75.7%
BFCL v3 (tool use)
Composite score6.86
Community ratingNo reviews yet

VRAM & disk per quantization

QuantVRAMDiskRAMContext
Q4_K_M266 GB264.5 GB10240K

Strengths & weaknesses

Strengths: Exceptional long context (RULER held to 1M, 4M-token inference); Strong general knowledge and instruction following (IFEval 89.1); Novel hybrid Lightning + Softmax attention MoE

Weaknesses: Non-reasoning; weaker GPQA (54.4) and competition math; Released Jan 2025, behind current models