Chinese startup MiniMax, known for its Hailuo AI, has released a new large language model, MiniMax-M1. It has 1 million input tokens (80,000 output), making it one of the largest AI models by this metric.
Chinese startup MiniMax, known for its Hailuo AI, has released a new large language model, MiniMax-M1. It has 1 million input tokens (80,000 output), making it one of the largest AI models by this metric.
By comparison, OpenAI’s GPT-4o has a context window of just 128,000 tokens—enough to exchange the volume of a single book between a user and a model in a single interaction. With 1 million tokens, MiniMax-M1 can exchange the volume of a small collection or series of books, VentureBeat notes. Google’s Gemini 2.5 Pro also offers a context window cap of 1 million tokens, and is reportedly working on a window of 2 million.
A «context window» in large language models (LLMs) refers to the maximum number of tokens that a model can process at one time. Tokens are basic units of text that can include whole words, parts of words, punctuation marks, or code characters. These tokens are converted into numeric vectors that the model uses to represent and manipulate values using its parameters (weights and biases).
M1 is available in the open source community of artificial intelligence code sharing Hugging Face and in the code sharing community GitHub under the Apache 2.0 license. This means that companies can use it for commercial purposes and modify it as they see fit without restrictions or payment.
M1 is trained using reinforcement learning using an innovative, inventive, and highly efficient technique. The model is trained using a hybrid Mixture-of-Experts (MoE) architecture with a lightning-fast attention mechanism designed to reduce inference costs.
According to benchmark data, MiniMax-M1 consumes only 25% of the floating point operations (FLOPs) required by DeepSeek R1 at a generation length of 100,000 tokens. M1 also competes with OpenAI o3, Gemini 2.5 Pro, Claude 4 Opus, DeepSeek R1, DeepSeek R1-0528, and Qwen3-235B in various benchmarks (AIME 2024, LiveCodeBench, SWE-bench Verified, Tau-bench, and MRCR), where it outperforms in some metrics and lags in others.
Day 1/5 of #MiniMaxWeek: We’re open-sourcing MiniMax-M1, our latest LLM — setting new standards in long-context reasoning.
— MiniMax (official) (@MiniMax__AI) June 16, 2025
— World’s longest context window: 1M-token input, 80k-token output
— State-of-the-art agent use among open-source models
— RL at unmatched efficiency:… pic.twitter.com/bGfDlZA54n
MiniMax has made it clear in its posts that it is trying to displace DeepSeek as China’s leading AI player. It was founded in late 2021 and is backed by investors including Alibaba and Tencent.
This winter, the startup launched three new AI models: MiniMax-Text-01 for text only, MiniMax-VL-01 for image and text recognition, and T2A-01-HD for generating sound, including speech. The developers claim that all of them are better than AI models from Google and Anthropic.
