LLM VRAM Calculator
Use llm vram calculator locally with explicit settings, interactive results and complete export.
Enabling local controls…
Interactive local workbench
Start explicitly to create a complete export. Editing clears stale results.
Model, sources and limits
Weights = totalParameters × bits /8; KV = 2 × layers × KVheads × headDim × context × batch × KVbytes. BinaryGiB, explicit overhead/runtime buffer. Quant labels use nominal bits, not exact GGUF size. MoE uses total stored parameters; no invented quality-retention percentages. Even GPU split is an assumption; no speed/fit/hardware guarantee. Reverse GPU fit uses the same assumptions and11 source-class GPU options. Apple unified-memory capacity scenarios use an editable72% assumption, full weights+KV+runtime regardless of offload; capacities are examples, not product availability or a fit guarantee. Each text field≤32KiB, complete result≤1MiB. Worker startup10s/process1s; edits cancel stale work.
How to use this tool
Review labeled settings and assumptions before starting. Everything stays in this tab; no AI requests, model execution, persistent storage or remote assets.
When not to use it
Weights = totalParameters × bits /8; KV = 2 × layers × KVheads × headDim × context × batch × KVbytes. BinaryGiB, explicit overhead/runtime buffer. Quant labels use nominal bits, not exact GGUF size. MoE uses total stored parameters; no invented quality-retention percentages. Even GPU split is an assumption; no speed/fit/hardware guarantee. Reverse GPU fit uses the same assumptions and11 source-class GPU options. Apple unified-memory capacity scenarios use an editable72% assumption, full weights+KV+runtime regardless of offload; capacities are examples, not product availability or a fit guarantee. Each text field≤32KiB, complete result≤1MiB. Worker startup10s/process1s; edits cancel stale work.