The MiniMax M2.7, an enhanced version of the M2.5 model, is now available, featuring open weights through NVIDIA and the open-source inference ecosystem. This sparse mixture-of-experts (MoE) model, optimized for efficiency, boasts a 230B-parameter capacity with a mere 10B active parameters at a 4.3% activation rate, ideal for various tasks in reasoning, ML research, software engineering, and office work. Utilizing advanced techniques such as Rotary Position Embeddings (RoPE) and Query-Key Root Mean Square Normalization (QK RMSNorm), it excels in complex agentic tasks while minimizing inference costs via a top-k expert routing mechanism. Developers can leverage NVIDIA NemoClaw for seamless deployment and running of OpenClaw assistants. Collaboration with the open-source community has resulted in performance enhancements via vLLM and SGLang frameworks, achieving substantial boosts in throughput. For fine-tuning, the NVIDIA NeMo Framework offers tools and recipes for effective integration. Explore MiniMax M2.7 on Hugging Face and NVIDIA’s platform for deployment options.
Source link
