Ornith 1.5 35B-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, top-8 routing per config.json). paramsB (35.95B) is the exact safetensors total from the Hugging Face API; activeParamsB (3B) is taken from the "A3B" designation in the official model name — Ornith's card does not state an exact activated-parameter figure in prose, so treat the 3B as the vendor's rounded label rather than a precise count. All 256 experts must be VRAM-resident, so the weights term uses the 35.95B total. maxContextLength 262,144 per the card (YaRN factor 4 extends to ~1M; not modelled here). Pulled 2026-09-19.