Kimi Linear 48B-A3B — VRAM requirements

Weights = params × bytes-per-parameter for the selected quantization.KV cache is always computed at FP16 and assumes batch size 1. A flat +15% is added on top for CUDA context/activation overhead — a rule of thumb, not a measured figure. Full methodology on the interactive calculator.

Sparse mixture-of-experts (256 experts + 1 shared). paramsB (49.12B) is the exact safetensors total from the Hugging Face API; activeParamsB (3B) and maxContextLength (1M) are both from Moonshot's own model-card table ("#Total Params 48B, #Activated Params 3B, Context Length 1M"). **config.json publishes no max_position_embeddings for this model**, so the 1M figure comes from the card, not from config. **This site's KV-cache term will substantially overestimate this model.** Kimi Linear is a hybrid architecture with a 3:1 ratio of Kimi Delta Attention (a linear/recurrent mechanism with fixed-size state) to global MLA layers; Moonshot states this cuts KV-cache requirements by up to 75% versus full attention. Neither linear attention nor MLA compression is modelled in src/lib/vram.ts, which assumes a standard per-head KV cache for all layers. Pulled 2026-09-19.