MiniMax drops open-source M1 reasoning model with 1M context

MiniMax-M1 is a new open-source reasoning model that pairs an industry-leading 1 million token context window with an 80,000-token reasoning output. If you need an AI to ingest massive codebases or hundreds of pages of documentation and work through complex, multi-step logical deduction, this model is built for that heavy lifting.

What it can do

MiniMax released two variants: MiniMax-M1-40k and MiniMax-M1-80k. On the coding benchmark SWE-bench, they achieved 55.6% and 56.0% respectively, landing right on the heels of DeepSeek-R1-0528 at 57.6% while easily surpassing other open-weight models.

Long-context processing is where M1 really shines. Backed by its 1M token window, the model ranked second globally on long-context tasks, trailing only Google's closed-source Gemini 2.5 Pro while outperforming OpenAI o3 and Claude 4 Opus.

On agent tool use via TAU-bench, MiniMax-M1-40k also led all open-weight models and beat Gemini 2.5 Pro. Benchmark data confirms that expanding the test-time thinking budget to 80,000 tokens consistently improves reasoning accuracy across difficult tasks.

How it cuts compute by 70%

Long-context reasoning is notorious for melting GPU clusters, but M1 solves this through two core engineering breakthroughs.

The first is a proprietary hybrid-attention mechanism powered by Lightning Attention. When running deep reasoning chains up to 80,000 tokens, M1 requires only about 30% of the computing power needed by DeepSeek R1, effectively slashing 70% of the inference compute overhead.

The second is CISPO, a custom reinforcement learning algorithm. Instead of conventional token updates, CISPO clips importance sampling weights to boost RL training throughput. In AIME experiments, it converged twice as fast as other RL algorithms, outpacing ByteDance's DAPO and outperforming the GRPO algorithm used in early DeepSeek training.

Thanks to these optimizations, the entire reinforcement learning phase required only 512 H800 GPUs running for three weeks. The compute rental bill came out to just $534,700, an order of magnitude lower than initially anticipated.

Where to run it

MiniMax has open-sourced the complete model weights and technical documentation on Hugging Face and GitHub.

For local hosting and inference, vLLM and Transformers already support M1 out of the box, with SGLang integration currently underway.

If you prefer not to manage infrastructure, MiniMax offers unlimited free access on its official web platform and mobile app.

For API users, pricing is very aggressive. Inputs under 200k tokens cost $0.40 per million input tokens and $2.20 per million output tokens, undercutting DeepSeek-R1. The 200k to 1M token tier runs at $1.30 per million input tokens and $2.20 for output, offering long-context reasoning capabilities that competing open APIs do not provide.

Why it matters

For developers building autonomous software engineering agents, document review pipelines, or complex workflows, M1 solves the painful trade-off between massive context and deep logical thinking. If you have been held back by context limits on open models, it is well worth testing immediately.

We think M1 proves that frontier reasoning does not require endless capital burn. By refining attention architecture and RL sampling math, MiniMax showed that a small cluster and roughly half a million dollars can train a model rivaling top closed-source systems, setting a fresh benchmark for open-source efficiency.