Introducing MiniMax-M1: An Innovative Open Source Model with 1 Million Token Context Capability

Chinese AI startup MiniMax has launched its open-source large language model, MiniMax-M1, under the Apache 2.0 license, allowing businesses to use and modify it freely. Known for its extensive context window of 1 million tokens—superior to competitors like OpenAI’s GPT-4—MiniMax-M1 excels in long-context reasoning and has been trained with a cost-effective reinforcement learning technique, making its total training expenses approximately $534,700. The model features a hybrid architecture with high-performance benchmarks: 86% on AIME 2024 and strong results in coding tasks, outpacing some open-weight models. It offers structured function calling capabilities, making it versatile for real-world applications. Recommended for deployment via vLLM or Transformers, MiniMax-M1 is positioned as a valuable resource for enterprises seeking advanced AI solutions without vendor lock-in, providing both operational efficiency and technical flexibility for teams managing AI systems. This release marks MiniMax’s commitment to scalable and accessible AI modeling.

Source link