MiniMax-M1 Revealed: Hybrid Attention, RL Training, and 1M Token Context

MiniMax Funding and Product Launch

MiniMax raised nearly $300 million in a new financing round, valuing the company at over $40 billion, and announced a series of products including the inference model MiniMax‑M1, video model Sea Shell 02, and the MiniMax Agent.

Technical Highlights from the M1 Closed‑Door Meeting

The M1 model is a 4.56 trillion‑parameter open‑source hybrid‑expert LLM that supports up to 1 million tokens of context and 80 k token inference output, matching the input length of closed‑source models such as Google Gemini 2.5 Pro.

Key innovations discussed:

Reinforcement‑learning (RL) training can endow limited‑context models with new capabilities and broaden their knowledge scope.

RL focused solely on mathematics and code tends to increase hallucinations.

Latent reasoning and visual‑latent reasoning are explored as ways for models to “think” with images.

Reward modeling, multi‑agent systems, and AI‑automation are identified as major RL challenges.

Hybrid linear attention dramatically reduces FLOPs per token, enabling efficient long‑context processing, though it may require more tokens to match full‑attention performance.

Long‑context handling is crucial for agents, allowing them to ingest entire codebases, API docs, and interaction histories in a single pass.

Hybrid architectures are projected to become mainstream, but current bottlenecks lie in infrastructure and GPU utilization.

System‑2 reasoning and self‑reflection emerge when models are given ample computation budget.

Performance and Cost

MiniMax‑M1 outperforms open‑source rivals such as DeepSeek‑R1 and Qwen3‑235B on complex software‑engineering and long‑context benchmarks. The RL training phase used 512 H800 GPUs for three weeks at a cost of $537,400, an order of magnitude cheaper than expected.

Implications for the LLM Industry

The advances demonstrate that efficient hybrid attention combined with targeted RL can achieve performance comparable to full‑attention models while keeping inference costs low, a direction likely to shape future large‑model research and deployment.