MiniMax unveils open source LLM with incredible 4M tokens

MiniMax is probably today better known here in the US as the Singaporean company behind it Hailuothe truth, the highest level of video AI competitors The way to run, OpenAI and Sora and Luma AI Dream Machine.

But the company has many innovations: today, for example, it announced the release and launch of the MiniMax-01 seriesa new family of models built to support long-term scenarios and advance AI development.

This list includes MiniMax-Text-01, a major language foundation (LLM), and MiniMax-VL-01multi-dimensional form.

Main information window

LLM, MiniMax-Text-o1, is very important for generating more than 4 million tokens in its window – equivalent to small library books. The knowledge window is the range of knowledge that an LLM can handle single exchange / outputand words and concepts represented as “symbols” of numbers, the LLM contained within the mathematics of the data that was taught.

And where Google led the pack with its version of the Gemini 1.5 Pro and 2 million token windowMiniMax somehow doubled!

Like MiniMax posted on his official X account today: “MiniMax-01 efficiently handles up to 4M signals – 20 to 32 times the amount of other advanced models. We believe that MiniMax-01 is about to support the increase in applications expected in the coming year, as agents need controllability and stable memory .

It is now available for download Hugging Face and Github under a MiniMax license customfor users to test directly Hailuo AI Chat (a ChatGPT/Gemini/Claude competition), and through MiniMax’s application programming interface (API)where third-party developers can connect their own software to them.

MiniMax offers a wide range of text editing APIs and models at competitive prices:

• $0.2 for 1 million tokens
• $1.1 per 1 million tokens issued

In comparison, OpenAI’s GPT-4o costs $2.50 per 1 million entry tokens through its API, more expensive 12.5 times.

MiniMax has also incorporated a Mixture of Experts (MoE) system with 32 experts to achieve scalability. This design is optimized for compute and memory while maintaining a competitive edge in benchmarks.

Creating a new environment with Lightning Attention Architecture

At the heart of the MiniMax-01 is the Lightning Attention system, an alternative to traditional Transformer hardware.

This design greatly reduces the computational complexity. The brand has 456 billion shares, of which 45.9 billion are controlled on the proposal.

Unlike previous constructions, the Lightning Attention uses a combination of traditional and SoftMax lines, making it more challenging for longer entries. SoftMaxfor those who are new to the concept like me, it is possible to change the input numbers to add up to 1, so that the LLM can estimate the meaning of the input.

MiniMax has also built its training and management systems to support the Lightning Attention architecture. Key changes include:

• MoE All-to-All Communication Optimization: Reduces inter-GPU connectivity.
• Varlen Ring Interest: Reduces computer waste for long-term maintenance.
• Using the Kernel Properly: Optimized CUDA strings improve Lightning’s sensitivity.

These advancements make the MiniMax-01 models available for real-world applications while keeping costs low.

Performance and Benchmarks

On popular text and multi-mode benchmarks, the MiniMax-01 competes with high-end models like GPT-4 and Claude-3.5, with strong results in long-term analysis. In particular, MiniMax-Text-01 achieved 100% accuracy on the Needle-In-A-Haystack task with a 4-million-token context.

These models also show a decrease in performance as the input length increases.

MiniMax regularly updates to expand the range’s capabilities, including multiple codes and modals.

The company sees the open source as a way to create basic AI capabilities in the design of AI assistants.

With 2025 predicted to be a breakthrough year for AI agents, the need for continuous memory and efficient communication between agents is growing. The new MiniMax was designed to meet these challenges.

Open to collaboration

MiniMax invites developers and researchers to explore what MiniMax-01 is capable of. In addition to opening, his team receives technical proposals and requests for cooperation on model@minimaxi.com.

With its commitment to cost-effective and scalable AI, MiniMax positions itself as a key player in the AI ​​assistant era. The MiniMax-01 series offers an exciting opportunity for developers to push the boundaries of what long-term AI can achieve.