Atlas

Models

← All models

MiniMax M1 40k

MiniMax · 2025-06-16 · 456.1B parameters

MiniMax M1 40K is MiniMax's open-weight hybrid-attention reasoning checkpoint with a 40K maximum generation length; MiniMax says it represents an intermediate phase of the 80K variant's training. The 456-billion-parameter MoE activates approximately 45.9 billion parameters per token, places one softmax-attention block after every seven Lightning Attention blocks, is described as natively supporting up to 1 million context tokens, and was trained with large-scale reinforcement learning using CISPO.

Benchmark scores

BenchmarkScore
AA-LCR51.7
AIME 202513.7
Artificial Analysis Intelligence Index14.4
Capability135.3
GPQA Diamond68.2
Humanity's Last Exam (Text-Only)7.5
IFBench (Artificial Analysis)41.2
LiveCodeBench (Artificial Analysis)65.7
MATH-500 (Artificial Analysis source)97.2
SciCode (Artificial Analysis)37.8
Terminal-Bench Hard2.3
τ²-bench (Telecom)31.6
Loading Atlas data…