October 23rd.MiniMax dilute technologyAnnounced todayOpen Source Mini Max M3.1 model weight for Flash, THE TOTAL PARAMETER 428B, ACTIVATE THE PARAMETER 23B。
The M3.1 Flash model capabilities were fully enhanced compared to the previous M3 model. In addition, M3.1 Flash further improves reasoning efficiency through the optimization of the full-model thin attention, quantitative awareness training and Power-aware learning:
- Full model thins attention:The first three layers of full attention are replaced by a dilution
- QAT:The router MoE expert moved to W4A4(except shared experts) and the diluted attention indexer and main branch moved from Q8KV8 to Q8KV4
- Power-aware learning:The fine distribution of the calculator power has kept the calculator continuously high and has resulted in a high degree of pressure on the hardware efficiency。
M3.1 Flash went online on October 19, and the real user processed more than 10 trillion Token within 72 hours。
Mini Max has declared CAC to be a platform of reasoning and cloud services, such as Zing, Hiranco, Kunlanc, Moor's Line, Wallstone, Sealight, Mute, Days, Weidar, AMD, and Together AI, Fireworks, Nebius, Baseten, Modal, DigitalOceanCo-push M3.1 Flash Open Source First Day Access.
In addition, the reasoning frameworks vLLM, SSGLang, resource FlagOS and Hugging Face, and the Demons and Open Source Community, together with MiniMax, will also promote the application of models in different chips, cloud platforms and development environments to continue to expand the global AI Open Source Ecology。
1AI ATTACHED OPEN SOURCE ADDRESSES AS FOLLOWS:
- Hugging Face: https://huggingface.co/MiniMaxi/MiniMax-M3.1-Flash
- Modified Community: https://modersscope.cn/models/MiniMax/MiniM Max-M3.1-Flash