MiniMax-Text-01

// Overview

MiniMax-Text-01 is the foundational text large language model in MiniMax's MiniMax-01 series, open-sourced on 15 January 2025 alongside the vision-language model MiniMax-VL-01. It is a Mixture-of-Experts model with 456 billion total parameters, of which 45.9 billion are activated per token across 32 experts (Top-2 routing) over 80 layers. MiniMax positions it as the first time linear ("lightning") attention has been scaled to a commercial-grade model of this size.

Its defining trait is a hybrid attention stack: within every block of 8 layers, 7 use lightning (linear) attention and 1 uses standard softmax attention. That mix keeps the cost of processing very long inputs close to linear, which is how MiniMax-Text-01 reaches a 1-million-token context during training and extrapolates to up to 4 million tokens at inference — among the longest contexts of any model at release. MiniMax reports 100% accuracy on a 4M-token Needle-In-A-Haystack retrieval test and a 0.910 RULER score at 1M tokens.

On core academic benchmarks MiniMax-Text-01 lands in the same range as GPT-4o and Claude 3.5 Sonnet — for example 88.5 on MMLU, 94.8 on GSM8K, and 89.1 on both IFEval and Arena-Hard. The weights are released openly under MiniMax's custom Model License Agreement on Hugging Face and GitHub (code is MIT), and MiniMax recommends vLLM for production serving. It is text-only; multimodal input lives in the separate MiniMax-VL-01. The later MiniMax-M1 reasoning model is built on this same Text-01 base.

// Benchmarks

1. MMLU88.5%
2. MMLU-Pro75.7%
3. GPQA Diamond54.4%
4. GSM8K94.8%
5. MATH77.4%
6. IFEval (avg)89.1%
7. Arena-Hard89.1%
8. C-SimpleQA67.4%
9. LongBench v2 (overall, with CoT)56.5%
10. RULER (1M tokens)0.91%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Pricing

MiniMax's official open-source announcement lists $0.20 per million input tokens and $1.10 per million output tokens; OpenRouter lists the same $0.20 in / $1.10 out for minimax/minimax-01. Weights are also free to download and self-host under MiniMax's Model License Agreement.

Pricing source ↗

// Strengths

• Extremely long context: 1M-token training window that extrapolates to up to 4M tokens at inference — among the longest available at release
• Lightning (linear) attention keeps long-context cost near-linear instead of quadratic, unlike a pure softmax transformer
• Strong long-context retrieval: 100% on a 4M-token Needle-In-A-Haystack test and 0.910 RULER at 1M tokens (per MiniMax)
• Competitive general benchmarks against closed frontier models of its era (MMLU 88.5, GSM8K 94.8, IFEval 89.1, Arena-Hard 89.1)
• Open weights with a low hosted API price ($0.20 in / $1.10 out per 1M tokens)
• Serves as the open base for the MiniMax-M1 reasoning model

// Best for

• Long-document and whole-codebase analysis that needs hundreds of thousands to millions of tokens of context
• Retrieval over very large inputs (long PDFs, transcripts, logs) where needle-in-a-haystack accuracy matters
• Self-hosted general-purpose chat and instruction-following where open weights are required
• A base model for fine-tuning or for building reasoning systems (as MiniMax did with M1)
• Cost-sensitive long-context API workloads via MiniMax or OpenRouter

// How to access