MiniMax-01: The Lightning-Attention Model Behind the Family

Long context is usually sold as a number on a spec sheet. The interesting question is what the model does internally to make that number affordable, because standard attention gets quadratically more expensive as the input grows. MiniMax-01, released in mid-January 2025, was an early large open-weight model to answer that question with a different attention mechanism rather than a cleverer cache.

It also became the base for MiniMax M1, the reasoning model that followed. If you want to understand why M1 handles long inputs the way it does, this is the place to start.

Two models under one name

MiniMax-01 is a series, not a single checkpoint. It consists of MiniMax-Text-01, the language model, and MiniMax-VL-01, a vision-language model built on top of it. Catalogue listings such as OpenRouter's minimax/minimax-01 expose the pair as one endpoint that accepts text and images and returns text.

The vision side follows a familiar design: a 303M-parameter vision transformer trained from scratch, a small two-layer projector, and Text-01 as the language backbone. Images are resized onto a grid between 336 by 336 and 2016 by 2016 pixels, split into patches, and encoded alongside a low-resolution thumbnail. MiniMax reports 96.4 on DocVQA and 91.7 on ChartQA, which points at documents and charts rather than general scene understanding as the strength. Its own table shows competitors ahead on some visual maths tasks.

How lightning attention changes the cost curve

Text-01 has 456B total parameters with 45.9B active per token, using a mixture-of-experts layer of 32 experts with top-2 routing. That part is conventional. The attention stack is not.

The model has 80 layers. Seven out of every eight use lightning attention, a linear-attention variant whose cost grows roughly in proportion to sequence length rather than with its square. Every eighth layer is ordinary softmax attention. The softmax layers keep the precise token-to-token retrieval that linear attention is weaker at; the lightning layers keep the overall cost of a very long input manageable.

The practical effect is that the model was trained with context extended to 1M tokens, and MiniMax says it can handle up to 4M tokens at inference. Hosted endpoints usually expose less: OpenRouter lists a context window of about 1M tokens at the time of writing.

What the long-context numbers say

MiniMax's headline long-context evidence is RULER, where it reports 0.963 at 4K tokens and 0.910 at 1M. A drop of about five points across a 250-fold increase in length is a good result, and it is the strongest argument for the hybrid design.

Two cautions. RULER is a synthetic retrieval-style test, and doing well at finding planted facts is not the same as reasoning well across a whole codebase or a year of contracts. And these are vendor-run figures. Context length vs effective context explains why the usable window is almost always smaller than the advertised one, and long context degradation describes how to measure where quality starts to fall on your own inputs.

On short-context general benchmarks, MiniMax reports MMLU 88.5, IFEval 89.1, HumanEval 86.9 and GPQA Diamond 54.4. Those were competitive in early 2025. They are not competitive with current models, and SimpleQA at 23.7 suggests factual recall is middling.

Licence and price

The weights are released under MiniMax's own Model Agreement, with the code under MIT. That is not the same as an MIT-licensed model. Read the agreement before you build a commercial product on self-hosted weights, paying particular attention to any use restrictions and attribution terms.

At the time of writing OpenRouter lists MiniMax-01 at $0.20 per million input tokens and $1.10 per million output. For long-document work that ratio is favourable, because input dominates the bill when you are sending hundreds of thousands of tokens and asking for a short answer.

MiniMax-01 or MiniMax M1

MiniMax M1, released in June 2025, is a reasoning model built on Text-01, with the same 456B total and 45.9B active parameters, trained further with large-scale reinforcement learning. The MiniMax M1 guide covers it in detail. The choice between them is mostly about whether you want the model to think before answering.

• Pick MiniMax-01 for long-document reading, extraction and summarisation where the answer is short and the reasoning shallow, and for anything involving images, since M1 is text-only.
• Pick M1 for maths, multi-step analysis, coding and tool use, where spending output tokens on reasoning buys accuracy.

For either, the newer generation is usually the better default today. The MiniMax family overview lays out how the line developed after these two.

A quick way to test it

Take one long document you already understand well, at least a few hundred thousand tokens. Write ten questions whose answers sit at different depths, including two that need facts from opposite ends of the text. Ask each question once with the full document and once with a retrieved excerpt. If the full-context answers are consistently better, MiniMax-01's window is doing real work for you. If they are not, retrieval with a cheaper short-context model will do the same job for less.