How much VRAM does MiMo-V2.5 310B (MoE) need at Q4_K_M? Xiaomi 1M SWA model

How this is calculated

The 310B parameter pool requires 174 GB of weights at Q4_K_M. The 48-layer architecture uses group-query attention with variable KV heads: 8 in the 9 full-attention layers, 4 in the 39 sliding-window layers. The QK and V dims share the same asymmetry as the Pro variant, so the effective head_dim is 160. Sliding-window attention caps 39 of 48 layers at a 128-token window. The key-value cache at the full 1M context uses about 48 GB. Activation and software overhead adds about 22 GB. In active-only mode the resident weights shrink to 8 GB, which drops total usage to 62 GB.

Verdict

The 244 GB resident total sums as 174 GB of Q4 weights, 48 GB of FP16 KV cache at the full 1M context (variable KV heads: 8 in the 9 global layers, 4 in the 39 sliding-window layers, asymmetric QK/V averaging to head_dim 160), and 22 GB of overhead. Hardware-wise that fits four H100 80GB cards or two H200 141GB cards with NVLink. Active-only loading shrinks the weights to about 8 GB by routing cold experts through RAM, dropping the total to roughly 62 GB, which fits a single 80 GB pro card or a 96 GB unified-memory workstation. The cost-performance case here is the practical one in the MiMo lineup: the Pro variant at 1.02T needs about 2.8x the resident VRAM and 2.8x the active parameters, while the 310B sibling carries the same architectural family at a footprint that fits single-datacenter-card hardware.