LLM VRAM calculator

Choose a model, the precision of its weights, your context length and how many requests run at once. The calculator adds the weights, the key-value cache and runtime headroom, and names the cheapest single GPU you can rent today that holds the total. DeepSeek R1 Distill Llama 70B at 4-bit with 4,096 tokens needs about 43.9 GB; the cheapest card that holds it is a 48 GB RTX A6000 at $0.35/hr (Thunder Compute; node size not published).

How the estimate is worked out

Weights = parameters × bits ÷ 8. For models with standard attention, the key-value cache = 2 × layers × KV heads × head dimension × context tokens × batch × 2 bytes. The total adds your headroom percentage to weights plus cache. Divide bytes by 1,000,000,000 for GB. Where a model's attention is not standard (sliding windows, compressed caches), only the weights are estimated and the calculator says so.

These are budget estimates, not measured peaks: the serving software, its own buffers and quantization format change the real figure, and fitting in memory says nothing about speed. For fine-tuning, open the model's own page, which adds LoRA and full fine-tune scenarios. Prices are the cheapest on-demand offers for one card, checked 12 October 2026; where a provider publishes no node size, the rate is per GPU and the result says so.

Not sure what your own card has? See how much VRAM do I have. All models: LLM VRAM requirements.