MiniMax-M1
Open weightsby MiniMax·China·Released
Open-weights 456B MoE with the largest free-tier context window in production: 1M tokens.
About this model
MiniMax-M1 (June 2025) is the open-weights flagship from MiniMax — the lab behind Hailuo video and the abab consumer chat series. The 456B-parameter MoE (45.9B active) is licensed under Apache 2.0 and ships with the largest production-ready context window in the open ecosystem: 1M tokens, fully usable (not degraded).
On agentic benchmarks MiniMax-M1 leads several Chinese open-weights peers — TAU-bench 73.4%, LiveCodeBench 65%, AIME 2024 86%. The combination of Apache 2.0 + 1M context + competitive quality is rare; it makes M1 the natural baseline for any open-weights research that needs very long context.
Strengths
- •1M-token context usable end-to-end (not degraded)
- •Apache 2.0 — most permissive license available
- •456B MoE with 45.9B active — efficient serving relative to scale
- •Strong agentic + coding benchmarks (TAU-bench, LiveCodeBench)
Limitations
- •MoE serving requires vLLM/SGLang stack
- •English chat quality trails Llama 3.3 70B on subjective tests
- •MiniMax less internationally known than DeepSeek/Qwen
When to use it
- →Whole-codebase or whole-corpus analysis (1M context)
- →Open-weights agentic workloads needing long-horizon reasoning
- →Research applications needing both Apache 2.0 + long context
Architecture & training
MiniMax-M1 paper (arXiv 2506.13585) introduces 'Lightning Attention' — a hybrid linear-attention variant that keeps memory linear in sequence length, enabling the 1M context at production cost. Trained on MiniMax's proprietary cluster.
Benchmarks
| Benchmark | Score | Bar |
|---|---|---|
| AIME 2024 | 86.0 | |
| TAU-bench | 73.4 | |
| LiveCodeBench | 65.0 |