MiniMax-M1-80k

by MiniMax

MiniMax M1 is described by its maker as the world's first open-weight, large-scale hybrid-attention reasoning model, featuring 456 billion total parameters with 45.9 billion activated per token. Released in June 2025 under the Apache 2.0 license, it pairs a 1M-token context window with a hybrid Mixture-of-Experts architecture and a Lightning Attention mechanism. The model scores 56.0% on SWE-bench (software engineering) and 86.0% on AIME 2024 (mathematics), and leads open-weight models on tool-use benchmarks such as TAU-bench. At 80K-token reasoning lengths, Lightning Attention uses roughly 30% of the compute that DeepSeek R1 requires, enabling efficient scaling of test-time compute across deep reasoning and long-context tasks. The 80K variant provides an 80,000-token reasoning budget for extended thinking. MiniMax trained M1 with CISPO, a reinforcement learning algorithm it reports converges about twice as fast as alternatives, completing training in three weeks on 512 H800 GPUs.

Key info

Input
Output
Features
Context window
1M
Max output
40K

Benchmarks

Independent benchmark scores: the Intelligence Index and the individual evaluations behind it, grouped by skill.

Global rank#254 of 520 LLMs
TierEfficient
Output speed52 tok/s
Intelligence Index (previous version)11.7
Coding
LiveCodeBench
71%
Agentic & knowledge work
Terminal-Bench Hard
3%
τ²-Bench Telecom
34%
Science & reasoning
Humanity's Last Exam
9%
GPQA Diamond
70%
Long-context reasoning
58%
Knowledge
MMLU-Pro
82%
Math & instruction following
AIME 2025
61%
IFBench
42%

Available routes

No routes currently available — MiniMax-M1-80k isn't routed through the Opper gateway right now. It may return.

Contact us about this model →

Available models from MiniMax

Start building with 700+ models

One API key for every major provider, up and running in minutes.

Get startedView Documentation