#BF16

'MiniMax-M2 is a compact MoE released with open weights on Hugging Face, optimized for coding and agentic tool use with about 10B active parameters per token and claims of ~2x speed at ~8% of Claude Sonnet cost.'

IBM released Granite 4.0, a hybrid Mamba-2/Transformer LLM family that cuts serving memory by over 70% for long-context inference while keeping strong instruction-following and tool-use performance.