VRAMNeeded Calculator

Methodology

This page is the exact rule every VRAMNeeded verdict uses: the calculator, model pages, /what-can-i-run/, the cheapest-way pages, Ask VRAMNeeded and Scan my GPU. The plain-language overview of where the data comes from is on Key facts.

The formula

Memory needed = weights + KV cache

  • Weights = the exact byte size of the published file at that precision (FP16/BF16 safetensors, Q8_0 GGUF, Q4_K_M GGUF; all shards added). Where a model ships a different 4-bit format (for example MXFP4 experts in gpt-oss, or Unsloth UD-Q4_K_XL when no Q4_K_M exists), the page says so.
  • KV cache = KV bytes per token × context tokens + any fixed part. Bytes per token = 2 × full-attention layers × KV heads × head dimension × 2 bytes (FP16 cache). Sliding-window or state layers that do not grow with context (Gemma, gpt-oss) are added as a fixed amount. Context is capped at the model's maximum.
  • 1 GB = 10⁹ bytes, the same unit GPU makers use for VRAM.

The 90% rule

  • Fits: weights + KV cache ≤ 90% of the card's memory. The other 10% covers the runtime, activations and the display.
  • Tight: only the 4-bit file fits, between 90% and 100% of memory. It may load but can run out of memory.
  • Won't fit: more than 100% at 4-bit.
  • We try 16-bit (or the model's native format, e.g. MXFP4 for gpt-oss), then 8-bit, then 4-bit and report the highest precision that fits. Every row names its precision, Mac rows included.

Macs (unified memory)

A Mac's memory is shared between the system and the GPU. We apply the same 90% rule to the full unified-memory size; macOS by default lets the GPU use a large share but not all of it, so treat "Tight" on a Mac as "probably won't load".

Multi-GPU and rentals

Rental prices are each provider's listed on-demand price on the checked date. Multi-GPU setups count total memory across cards and total $/hr; they need software that splits the model across cards. Cheapest single-card and cheapest overall are shown separately.

Sources and dates

Every number has a source link on its page and a checked date. If we can't source a number, the cell says "not yet sourced" — we don't estimate it. We do not show GPU purchase prices because we can't check them daily.

Catalog version

The model, GPU and price data together form a catalog version: cat-2026-11-03-46d273a7. It changes whenever any number changes, and dated changes are listed in the changelog. Our build fails if the calculator, the browser fit code and the published pages disagree on any of our golden test cases.

Golden tests

Before every publish we check fixed cases (for example Llama 3.1 8B at 4-bit, 8K context on an 8 GB RTX 4060; Gemma 4 31B with its fixed sliding-window cache; gpt-oss-120b with MXFP4 experts on a 96 GB Mac; context scaling from 8K to 128K) in both the Python build and the browser code. If any answer differs, nothing is published.

The answer is never sponsored

The fit verdict, the numbers and the order of results are never paid for. Affiliate and referral links appear only in clearly labeled rental, price and buy sections. See Disclosure.

Spotted a wrong number? Report an error.

Last checked: