How much VRAM does this model need?

What this estimate assumes

• Memory needed is the published weight file plus the KV cache. 1 GB = 1,000,000,000 bytes.
• KV cache is FP16. The fit module adds per-token KV bytes, the tokens actually used, and any fixed part stored on the model. Weight quantization does not change KV precision. This calculator does not offer Q8 or Q4 KV.
• A comfortable fit means the total stays at or under the catalog headroom (90% in this file). The other 10% stands in for the runtime, activations, and the display. That reserve is an assumption. Nothing is measured by running the model.
• Physical capacity is 100% of the card’s stated memory. A total above 90% and at or under 100% is a tight fit: it may load and it is not a comfortable fit. Over 100% will not fit.
• The highest precision that fits comfortably is chosen from the sourced files only: the native format when one is named (MXFP4 for gpt-oss), otherwise 16-bit, then 8-bit, then 4-bit. A missing file is skipped. It is not filled in with a guess.
• Context is capped at the model maximum when this catalog records one. Rows with no maximum keep the calculator’s top preset, 128K tokens, and the page says so. Presets above a recorded maximum are disabled.
• On a Mac the same 90% rule is applied to the full unified-memory size. macOS keeps memory for itself, so a tight Mac result often will not load.
• No inference is run here, and nothing you type is uploaded.

Open VRAMNeeded Sources Read the methodology