M1 New Open-Source AI Model From China SHOCKS The Industry (CRUSHES DeepSeek)
AI Summary
The video presents Miniax’s newly launched M1, a 1 million token language model accessible for free with no paywalls or API limits. M1 stands out with its exceptionally large context window—capable of handling a million input tokens and 80,000 token responses—which allows it to keep entire book series in short-term memory and produce long, coherent outputs. The model uses a mixture of experts architecture, activating only a portion of its 456 billion parameters per token, and implements “lightning attention,” a linear attention mechanism that keeps computational costs low even with very long sequences. This design enables much greater token efficiency and much lower training costs — around $534,700 compared to millions spent on comparable models like GPT-4. Reinforcement learning with a novel CISPO algorithm improves training stability and creativity. The training process incorporates a large dataset focused on STEM, code, and logic puzzles, progressing through increasingly complex tasks, with careful tuning to avoid issues like repetitive loops and numerical instability. M1 delivers competitive accuracy on benchmarks such as mathematical reasoning, code tasks, knowledge and logic, and tool use, though there remains room for improvement in real-world reasoning complexity. The model’s ability to generate long answers correlates with better performance. The open-source nature and permissive license allow deployment without vendor lock-in, making it attractive for privacy-focused use cases. The repository includes components for building an AI assistant, with tools for function calling and multimodal input/output integration.