AI CURATED HOT STORY · 10/23/2026
MiniMax has officially open-sourced the weights for its M3.1 Flash model, featuring a sparse architecture with 428 billion total parameters but only 23 billion active per inference. This design maintains massive knowledge capacity while significantly reducing computational load and memory usage.
Cross-verified reports highlight key innovations including full-model sparse attention, Quantization-Aware Training (QAT), and power-aware scheduling. vLLM announced Day-0 support on NVIDIA Blackwell/Rubin and AMD MI350 GPUs, utilizing NVFP4 KV caching and speculative decoding with DSpark draft models. These optimizations boost indexer decoding speed by up to 1.7x, enhancing throughput for long-context tasks.
With immediate ecosystem support from Huawei Ascend, NVIDIA, and major cloud platforms, M3.1 Flash demonstrates strong hardware-software co-design. It offers developers a path to deploy high-performance LLMs under constrained compute resources, potentially reshaping the cost-performance balance in the open-source AI landscape.
The line compares only sources observed throughout; gap hours are interpolated to keep the trend continuous.