VRAM explorer: which model fits your hardware
Pick an open model and see how much memory it needs and which machines it fits. No sign-up: the maths runs in your browser.
Work out the memory a model needs
Pick a model, a quantisation and a context. The bars show which machines it fits.
From a card we take off 0.5 GB. From unified or system memory, what the operating system keeps: 3 GB up to 16 GB, 4 GB up to 32 GB and 6 GB above.
Estimated memory needed
Weights = parameters × bits per weight ÷ 8 (Q4_0 4.65 · Q4_K_M 4.89 · Q8_0 8.5 · FP16 16). Cache = layers × KV heads × head dimension × 4 bytes, for each token of context and each slot. Compute buffers: an allowance, not a measurement, of 0.5 GB at one slot for models up to 24B and 1.1 GB otherwise; in our load logs they took 112 to 912 MiB on the GPU.
Checked on 2026-11-15 against 13 real loads on our two GB10 machines (Llama 3.1 8B, Qwen2.5 7B and Mistral Small 24B, 4k to 32k of context, at 1 and at 8 slots). The cache matches the loader's log to the MiB. The total comes out 0.4 to 1.5 GB above the size Ollama reported, and never below it: the gap is the part of the file Ollama keeps off the GPU, which ollama ps does not count, plus what is left of the allowance. From each machine we also take off what the operating system keeps.
Estimated speed: –
Which machines it fits
From least to most memory. Open a machine for its spec sheet and where to buy it.
Cloud GPU rental (hourly)
Tip:
Shall we set it up on your machine?
Get this analysis by email
Analysis saved. We will email it to you.
Next step: Hardware selector ROI calculator Grants
Checked against real loads
What this explorer works out, next to what Ollama reported for 13 loads on our two GB10 machines (128 GB of unified memory). The table is generated when the page is published, from the same sum as above.
| Load | Context × slots | Explorer | Ollama reported | Difference |
|---|---|---|---|---|
| Qwen2.5 7B Q4_K_M | 4k × 1 | 5.39 GB | 4.75 GB | +0.64 GB |
| Qwen2.5 7B Q8_0 | 4k × 1 | 8.83 GB | 7.89 GB | +0.94 GB |
| Qwen2.5 7B FP16 | 4k × 1 | 15.97 GB | 14.52 GB | +1.45 GB |
| Qwen2.5 7B Q4_K_M | 32k × 1 | 7.04 GB | 6.59 GB | +0.45 GB |
| Qwen2.5 7B Q8_0 | 32k × 1 | 10.48 GB | 9.74 GB | +0.74 GB |
| Qwen2.5 7B FP16 | 32k × 1 | 17.62 GB | 16.36 GB | +1.26 GB |
| Llama 3.1 8B Q4_K_M | 4k × 1 | 5.95 GB | 5.27 GB | +0.68 GB |
| Llama 3.1 8B Q4_K_M | 32k × 1 | 9.70 GB | 9.21 GB | +0.49 GB |
| Llama 3.1 8B Q4_K_M | 4k × 8 | 10.30 GB | 9.15 GB | +1.15 GB |
| Llama 3.1 8B Q4_K_M | 16k × 8 | 23.19 GB | 22.32 GB | +0.87 GB |
| Llama 3.1 8B Q4_K_M | 32k × 8 | 40.37 GB | 39.27 GB | +1.10 GB |
| Mistral Small 24B Q4_K_M | 4k × 8 | 20.88 GB | 19.79 GB | +1.09 GB |
| Mistral Small 24B Q4_K_M | 8k × 8 | 26.24 GB | 25.64 GB | +0.60 GB |
In all 13 loads the explorer comes out above what was measured, by at most 1.5 GB over what Ollama reported and 1.0 GB over the sum of every buffer in the log. Ollama reports the GPU buffers only; the explorer counts the whole file, because on unified memory it all comes out of the same pool.
What it does not cover: one hardware family and one engine (llama.cpp under Ollama 0.30.10 and 0.34.0), models of 7 to 24B, no partial CPU offload and no quantised cache. Models that mix layer types (Gemma 4, Qwen 3.6) do not follow this sum. How we checked it · Raw data (JSON)
Edge AI Hardware Selector
Answer 3 questions and we recommend the ideal hardware for your case.
1. What is your primary use case?
Write the same sum yourself in the Dev Playground (12 min) →
Then: which model fits your memory · the VRAM maths, step by step · how many tokens your text takes
Tell us what you want to run
Tell us what you want to run and on what budget. We will tell you which hardware you need, which model fits, and what to expect from it, before you spend anything.
102 free guides · 17 compliance templates