A lawyer pastes an entire 800-page contract into a model and asks it to trace every clause that contradicts a later amendment. The model reads the whole document in one pass, no chunking, no lost context.
That single-shot long read is the headline trick of MiniMax-M1 — and the part that is genuinely hard to copy.
The bet: long context without a compute blow-up
MiniMax-M1 was released by MiniMax (稀宇科技), a Shanghai lab, in mid-June 2025, and open-sourced under the Apache-2.0 license. MiniMax calls it the first open-weight, large-scale hybrid-attention reasoning model.
The architecture is a mixture-of-experts (混合专家, MoE) system:
- 456 billion total parameters, about 45.9 billion active per token.
- 32 experts with top-2 routing.
- A hybrid of Lightning Attention (a linear-attention mechanism) and standard softmax attention.
The hybrid is the point. Pure linear attention scales cheaply with length but can lose recall; pure softmax attention recalls well but scales badly. MiniMax's design keeps one softmax layer every eight layers so the model remembers, while the linear layers carry the long-range load. The payoff is a 1-million-token context window — eight times DeepSeek-R1's — with an 80,000-token reasoning output.
The number everyone quotes: $534,700
The detail that drew engineers' attention was the training cost. MiniMax says the entire reinforcement-learning phase ran on 512 NVIDIA H800 GPUs for three weeks at a rental cost of about US$534,700 — an order of magnitude below what many assumed a frontier-scale RL run would require.
Two innovations drove that:
- Lightning Attention cut the compute needed for long-context reasoning. MiniMax reports that generating 80,000 tokens of reasoning needs only about 30% of the compute of DeepSeek-R1.
- CISPO, a reinforcement-learning algorithm (Clipped IS-weight Policy Optimization) that clips importance-sampling weights rather than full policy updates. MiniMax says it converged about twice as fast as alternatives such as ByteDance's DAPO on AIME math tests.
The technical report, "MiniMax-M1: Scaling Test-Time Compute Efficiently with Lightning Attention," lays out the method. The efficiency story is credible because it is architectural, not just a smaller model.
What it is good at
MiniMax evaluated M1 on 17 benchmark sets and reported:
- SWE-bench Verified: 55.6% (40k thinking budget) and 56.0% (80k) — slightly below DeepSeek-R1-0528's 57.6% but above other open-weight models at the time.
- Long-context understanding: ranked second globally behind Gemini 2.5 Pro, ahead of OpenAI o3 and Claude 4 Opus, per MiniMax's runs.
- τ-bench (agent tool use): led all open-weight models and beat Gemini 2.5 Pro in MiniMax's evaluation.
The pattern is consistent: M1 is strongest where length and tool use matter — reading whole codebases, whole books, or long multi-turn agent sessions — and merely competitive on short math sprints.
Access and price
M1 ships as two checkpoints, M1-40k and M1-80k (different thinking budgets), on Hugging Face (MiniMaxAI/MiniMax-M1-80k) and GitHub under Apache-2.0. Inference is supported by vLLM and SGLang.
On MiniMax's API the launch pricing was:
- US$0.40 per million input tokens and US$2.20 per million output tokens for inputs up to 200K.
- US$1.30 per million input tokens for the 200K–1M tier (output unchanged) — a tier DeepSeek-R1 did not even support.
- Unlimited free use in the MiniMax app and web interface.
In RMB terms that is roughly 0.8 to 24 yuan per million tokens depending on length.
The catch: you still need the memory
Efficiency at training and per-token inference does not mean a small model. All 456 billion parameters must sit in GPU memory. Community estimates put a full-precision deployment near 912 GB of VRAM, or about 228 GB quantized to 4-bit — still a multi-GPU, datacenter-class machine. The "cheap" claim is about the math of long context, not the size of the box.
Honest limitations
- Benchmarks are vendor-reported. The SWE-bench and τ-bench numbers come from MiniMax's own runs and scaffolding; independent reproduction is limited, and methodology notes show the SWE-bench result used a 486-task subset rather than the full 500.
- Still behind R1 on coding. At 56.0% versus 57.6% on SWE-bench Verified, M1 trails DeepSeek-R1-0528 on the most-watched coding test.
- Huge to self-host. 456B parameters mean datacenter hardware regardless of attention efficiency.
- Long context is the differentiator, not raw reasoning. On short reasoning sprints, several open and closed models score higher; M1 wins on length and agent loops.
- It has been succeeded. MiniMax has since shipped the M2 family, which is smaller and cheaper to serve; M1 remains the lab's long-context flagship but not its newest model.
What readers can do now
- Use the free tier for long-document work. Paste a whole book, contract, or codebase into the MiniMax app or API and test the 1M-token window on a task that breaks shorter-context models.
- Pull the weights for self-hosting. Grab MiniMaxAI/MiniMax-M1-80k from Hugging Face and serve it through vLLM or SGLang on a multi-GPU box; budget for the 228 GB+ quantized memory footprint.
- Benchmark the cost claim yourself. Run an 80K-token reasoning task on M1 and a comparable model and compare GPU-hours — the efficiency thesis is the one worth verifying, not the leaderboard scores.