MiniMax goes open-source with 4M-token context models

MiniMax just open-sourced the MiniMax-01 series, a pair of heavyweight models built for massive inputs: the MiniMax-Text-01 language model and the MiniMax-VL-01 vision-language model. Their main flex is an unprecedented 4 million tokens of context window, roughly 20 to 32 times longer than standard frontier models.

That scale lets you dump hundreds of source files, entire codebases, or dozens of full-length books into a single prompt without chunking or losing continuity.

How it scales to 4 million tokens

Standard Transformer models hit a computational wall with massive context because attention costs scale quadratically. MiniMax worked around this by implementing Lightning Attention, a linear attention architecture scaled up for commercial-grade weights.

Inside the model, every 8-layer block uses 7 layers of linear Lightning Attention and only 1 layer of traditional SoftMax attention. That hybrid design keeps attention costs close to linear while preserving dense reasoning.

Under the hood, it is a 456-billion-parameter MoE architecture that activates 45.9 billion parameters per token. In tests, the model scored 100% accuracy on a 4-million-token needle-in-a-haystack retrieval benchmark with minimal degradation as inputs grew.

Availability and how to run it

MiniMax released the full weights for both MiniMax-Text-01 and MiniMax-VL-01 on GitHub. Because hosting a 456B MoE model requires a serious enterprise cluster, local deployment is strictly for teams with heavy hardware.

For everyday testing, you can use both models directly inside MiniMax Chat on the web. Developers can also connect via API, priced at $0.20 per million input tokens and $1.10 per million output tokens.

Why it matters

If you build autonomous agents or deep research tools that choke on token limits, MiniMax-01 gives open-source builders an alternative to closed-source million-token APIs. Testing it makes sense if your workflow depends on massive context dumps without losing long-range reasoning.

More broadly, this is the first time linear attention has successfully scaled inside a commercial-grade 400B-plus model. It proves you do not need pure brute-force compute to unlock multi-million context, which could shift how future open models tackle sequence length.