Atlas

Models

← All models

MiniMax M1 80k

MiniMax · 2025-06-16 · 456.1B parameters

MiniMax M1 80K is the final long-thinking checkpoint in MiniMax’s two-version M1 release, an open-weight hybrid Lightning/softmax-attention MoE reasoning model built from MiniMax-Text-01. MiniMax reports 456 billion total and 45.9 billion activated parameters, and trained this checkpoint by extending the 40K model through staged RL output-length limits of 48K, 56K, 64K, 72K, and 80K using CISPO; the released safetensors contain exactly 456,089,655,296 parameters.

Benchmark scores

BenchmarkScore
AA-LCR54.3
AA-Omniscience Index-47.4
AIME 202561.0
Artificial Analysis Intelligence Index17.7
Artificial Analysis Omniscience Accuracy21.1
Artificial Analysis Omniscience Hallucination Rate86.8
Capability139.0
CritPt0.0
Fiction.liveBench69.4
GPQA Diamond69.7
Humanity's Last Exam (Text-Only)8.2
IFBench (Artificial Analysis)41.8
LiveCodeBench (Artificial Analysis)71.1
MATH-500 (Artificial Analysis source)98.0
MMLU Pro81.6
SciCode (Artificial Analysis)37.4
Terminal-Bench Hard3.0
τ²-bench (Telecom)34.2
Loading Atlas data…