MiniMax just released two open source AI models that set a new standard for long-form content processing. The MiniMax-Text-01 language model and MiniMax-VL-01 visual model can now handle up to 4 million tokens – that’s 20-32 times more context than leading models like GPT-4 and Claude.
The technical specs are impressive. MiniMax-Text-01 packs 456 billion total parameters with 45.9 billion activated per token. It uses a hybrid approach combining Lightning Attention, Softmax Attention, and Mixture of Experts. Seven out of eight layers use linear attention, with traditional SoftMax in the remaining layer.
On the visual side, MiniMax-VL-01 was trained on 512 billion vision-language tokens. Early testing shows strong performance on multimodal understanding tasks.
Both models are available now on GitHub and Hugging Face, so you can run them locally for free. This move puts serious pressure on closed source AI companies by making long-context models openly accessible.
The 4M token context window fundamentally changes what’s possible with AI text processing. Previous models struggled with long documents, but MiniMax-Text-01 maintains consistent performance even with massive inputs.
You can check out the models here:
GitHub: https://lnkd.in/gFj3JM_H
Hugging Face: https://lnkd.in/gFj3JM_H
Announcement: https://lnkd.in/gpeN3rdm
This release is part of a broader trend toward open source AI that I covered in my post about Mistral’s Codestral model: https://adam.holter.com/codestral-25-01-mistrals-new-code-model-makes-top-ai-assistants-free/
As more companies open source their models, we’ll see faster progress in AI capabilities. The real innovation happens when developers can freely experiment with and build upon each other’s work.
