MiniMax's New Open-Source Model: Q4 Weights Fit 24GB GPUs, 512K KV Cache Needs ~104GB

Tests of MiniMax's new open-source model show BF16 weights around 66GB, while Q4KM quantization drops to about 20.8GB, fitting 24GB GPUs. However, serving 512K-token contexts requires roughly 104GB of KV cache memory.

2026-10-26 ~ 2026-10-26 · 2 related posts

• New model weight sizes: Q4_K_M GGUF fits a 24GB card with ~20.8GB — AIFlow_ML · 2026-10-26
• Community math: MiniMax model's 512K context needs ~104GB KV cache, multi-GPU only — AIFlow_ML · 2026-10-26