Every long document a model reads comes with a hidden tax. The further it looks back, the more compute it burns, until most models quietly cap out around 128,000 tokens and call it a context window.
MiniMax-01 refused that cap. Its bet was architectural: change the part of the model that decides what to remember, and the length problem starts to dissolve.
Two models, one architecture
On January 15, 2025, Shanghai-based MiniMax released and open-sourced the 01 series: MiniMax-Text-01, a text foundation model, and MiniMax-VL-01, a visual multimodal (多模态) model that can also read images. Both share the same core design.
The headline number is context. MiniMax-01 was trained at a one-million-token context and can process up to four million tokens at inference — the company positioned that as 32 times the context of GPT-4o and 20 times that of Claude 3.5 Sonnet. For reference, four million tokens is roughly the length of dozens of long books in a single prompt.
The actual innovation: Lightning Attention
The trick is not more memory, but a different attention mechanism. Traditional Transformers use softmax attention, whose cost grows with the square of the input length — that quadratic curve is what caps context windows. MiniMax-01 instead uses "Lightning Attention," a linear-attention mechanism whose cost grows roughly linearly with length.
The design is hybrid. In every block of eight layers, seven use Lightning Attention and one uses standard softmax attention. MiniMax described this as the first time linear attention was scaled to a commercial-grade model, which forced a near-total rebuild of its training and inference systems, including MoE communication and long-sequence kernels.
The model itself is large: 456 billion total parameters, with 45.9 billion activated per token in a mixture-of-experts layout. On standard academic benchmarks the company reported performance comparable to GPT-4o and Claude 3.5 Sonnet.
The reason the hybrid 7-to-1 split matters is that pure linear attention is cheap but can lose fine detail, while pure softmax attention is precise but quadratic. By keeping one softmax layer in every eight, MiniMax lets the model periodically re-ground itself against the exact relationships in the input, so the cheap linear layers do most of the long-range scanning without the whole system drifting. For a developer, the takeaway is that the four-million-token window is not a marketing trick bolted onto a normal model — it required rebuilding the training and inference stack, which is why the company frames it as a research bet as much as a product.
Why this points at agents
MiniMax framed the long context explicitly as agent infrastructure. A single agent with persistent memory, or many agents talking to each other, needs to hold far more state than a chat session ever does. A four-million-token window means an agent can keep a huge working set — documents, logs, prior steps — in view at once.
The pricing was stated in RMB: about 1 yuan per million input tokens and 8 yuan per million output tokens, or roughly $0.14 / HK$0.78 input and $1.10 / HK$6.20 output at current rates. That is the kind of unit economy that makes long-context agents cheap enough to actually run.
What it is good at, and not
On long-context evaluations MiniMax reported a 100% score on a four-million-token "needle in a haystack" retrieval test, and the smallest performance drop as input length grew — exactly the behavior you want from linear attention. The multimodal (多模态) VL-01 model showed strong results on vision-understanding sets as well.
The catch is that raw context length is a means, not a product. Having room for four million tokens does not guarantee the model uses them wisely, and most real workflows do not yet know how to feed or manage that much state. The architecture is the story; the applications are still catching up.
Why it matters
Long context has always been the quiet bottleneck of practical AI: most real work arrives as sprawling, messy documents, not tidy chat turns, and the old ~128K cap forced teams to chop that material into pieces and stitch the answers back together. MiniMax-01's bet is that removing the cap changes what kinds of applications are even worth attempting — an agent that holds an entire repository or a year of logs in view can reason across them instead of forgetting the start by the end. The catch is that the architecture is ahead of the habits around it; the value shows up only once workflows learn how to feed and use that much state. Still, the direction points at agents that remember, which is the part that has always been hardest.
Key takeaways
- MiniMax-01 pairs MiniMax-Text-01 with a multimodal (多模态) MiniMax-VL-01, both open-source.
- Lightning Attention (linear) scales context to 4M tokens at inference vs. the usual ~128K cap.
- The model is 456B total / 45.9B active (MoE), with 7 linear + 1 softmax attention layers per 8.
- Pricing is ~1 RMB in / 8 RMB out per million tokens (~$0.14 / ~$1.10), aimed at agent workloads.
What readers can do now
- Test long context: throw an entire codebase or a stack of PDFs at MiniMax-Text-01 to see retrieval across millions of tokens.
- Build a memory-heavy agent: use the 4M window as persistent state for a multi-step agent instead of external vector stores.
- Use the open weights: pull MiniMax-Text-01/VL-01 to experiment with linear-attention long-context research on your own hardware.
- Structure the input deliberately: a four-million-token window rewards well-organized prompts — feed documents with clear headers and section markers so the model knows where to look, instead of dumping raw text and expecting automatic understanding.
Honest limitations
Four-million-token context is only as useful as your ability to structure that much input; naive prompts will not automatically benefit. The architecture is novel and less battle-tested than standard Transformers, so expect rough edges in tooling and quantization. Benchmark gains are company-reported, and "32x GPT-4o" compares a specific context figure, not overall quality. The open weights help researchers, but running 456B-class models still demands serious hardware. Treat MiniMax-01 as a promising architectural bet, not a finished replacement for established models.