Will it fit?

Weights plus KV cache plus stated overhead, checked against the live book — the working shown, every approximation named. The fit is deterministic arithmetic; the speed is yours to measure. Optional: give your measured tokens/s and each machine prices itself per million tokens — we do the division, you own the throughput claim.

model Llama 3.1 8B Llama 3.1 70B Llama 3.1 405B Mistral 7B Mixtral 8x7B Qwen2.5 7B Qwen2.5 32B Qwen2.5 72B Gemma 2 9B Gemma 2 27B custom size…

precision FP16INT8INT4

context 4k8k32k128k

batch

class anyon-demandinterruptible

your tok/s (opt)

Spec-sheet arithmetic, not a benchmark: VRAM need is public math (the assumptions above are the whole list), the book is the live tape, and we never invent a tokens/s number — that one is yours. Agents: GET /api/fit or the will_it_fit MCP tool, docs.