How our estimates and guides are produced

Our calculator produces planning estimates for one conversation with the complete model loaded on a GPU. It is not a measurement of your computer or a promise that a runtime will support it.

Methodology updated 16 September 2026. Site owner: Zoe Stimpson.

Where the inputs come from

Model architecture inputs include layer count, attention KV heads, head dimension and a recorded context limit. Each model page links to its architecture source. Most weight sizes are estimated from parameter count and effective bits per weight; gpt-oss uses rounded published MXFP4 download sizes including all experts. Active parameters alone do not describe the memory needed to hold an MoE model.

The calculation

  1. Convert weight size from decimal GB to GiB by multiplying by 1,000,000,000 and dividing by 1,024³.
  2. Estimate FP16 KV cache in bytes as 2 × layers × KV heads × head dimension × context tokens × 2. The first factor accounts for keys and values; the last is bytes per FP16 value.
  3. Add a runtime allowance equal to the larger of 1 GiB or 10% of weight memory.
  4. Divide the total by 0.9, leaving 10% of the GPU target as headroom.
  5. For system RAM, add 8 GiB to weights and round up to a multiple of 8 GiB.
  6. For Mac unified memory, take the larger of GPU target plus 8 GiB or GPU target divided by 0.75, then round up to 8 GiB.

A worked example

Llama 3.1 8B at Q4_K_M and 4,096 tokens has estimated weights of 4.52 GiB, cache of 0.50 GiB and runtime allowance of 1.00 GiB. Including headroom gives a 6.69 GiB GPU target, 16 GiB system RAM target and 16 GiB unified-memory target. Rounding the display does not change the full-precision fit check.

What the estimate leaves out

We do not predict CPU-offload speed, multi-GPU splitting, image processing, concurrent requests or every runtime allocation. Gemma and gpt-oss estimates use full-context cache per layer without sliding-window savings. Compressed cache and runtime-specific optimisations can reduce memory; temporary allocations or additional features can increase it.

Runtime support is deliberately limited: vLLM estimates use FP16 or native gpt-oss MXFP4 and exclude Mac mode. A capacity match still requires checking the runtime’s supported GPU, drivers and model format. System RAM allowances are planning rules, not experimentally measured peaks.

Editorial and source policy

Guides connect documented runtime behaviour with worked examples calculated by our own tool. We prefer model publishers and runtime documentation for technical facts, link readers to the relevant sources and identify dated checks where available. AI-assisted drafting and development are used on this site; this is not a substitute for source checks or hands-on testing. We do not invent test results, reviewer credentials or personal experience.

Our model pages share a consistent format so assumptions are easy to compare. Model-specific inputs and resulting estimates differ. The catalogue is not a performance ranking. Results remain alphabetical, and affiliate commission does not alter the calculations.

Measured results and corrections

Benchmarks are labelled separately with their workload and software version where supplied. “Source checked” means we checked the published result, not that we reproduced it. Incomplete test environments and differences between software versions are disclosed alongside results. Guides and data can become outdated; no continuous verification or fixed update frequency is implied.

When reporting a discrepancy, include the page URL, exact model file, runtime version, context, available memory and a public source or redacted measurement. Distinguish download size from peak memory and prompt-processing speed from generation speed. Material corrections should explain which input or assumption changed, rather than imply historical results were newly measured.

Commercial links

Provider links can earn a commission, as disclosed near hardware recommendations. Links do not constitute a guarantee of availability, price or compatibility. Sponsored material, if published, must be labelled. Advertising will remain separate from calculator controls and download actions.