VRAMNeeded Calculator
NEW

Skip the GPU: run open models through Novita AI's model APIs, or give your agents a Novita Sandbox. New categories: AI APIs · Sandboxes → Novita AI referral links: we may earn from its links. Disclosure.

Model in, VRAM out.

How much VRAM do you actually need?

We earn commissions when you shop through the links below.

Cheapest way to run it

All models: own GPU vs rent vs API →

Weights + KV cache from real Hugging Face file sizes, 90% fit rule, every number sourced.

How we calculate + sources ⓘ
  • Weights: the GGUF file sizes listed on Hugging Face.
  • KV cache: worked out from each model's config.json (FP16 cache).
  • Fit rule: weights + KV cache must stay under 90% of your GPU memory.
  • Prices: copied from each provider's page, with a Verified time on every row.

Full methodology and sources →

Get price drop alert for this model

Pick a model and your GPU, or enter your GB to see what fits. Every number links to its source.

Can my GPU run it?

Can a 4090 run gpt-oss-120b?Can a 4090 run Nanbeige4.3 33B?Can a 3060 run Qwen3 14B?Can a 5090 run Gemma 4 31B?Can a 5070 run gpt-oss-20b?Can a 3090 run Qwen3.8 27B?

4-bit Q4_K_M weights + 8K KV cache under 90% of VRAM, the same rule as our calculator.

VRAM calculator

Pick a workload, a model and your graphics card. The answer updates as you choose.

Estimates. LLM numbers use real download sizes from Hugging Face plus the KV cache, the same as our VRAM table and model pages. Checked 2026-11-03.

How it works in 10 seconds

  1. 1
    Pick a model

    Llama, Qwen, Gemma, DeepSeek and more, or just type your GPU memory.

  2. 2
    See the VRAM

    Weights plus KV cache at your quantization and context length, with sources.

  3. 3
    Run it

    Fits on your card? Great. If not, compare rental prices or cards that fit.

Too big for your card? Rent the VRAM by the hour, or upgrade.

Run it on a cloud GPU

No upgrade needed

Rent a 24–80 GB GPU for the hour you need it.

Vast.ai Referral link Runpod Referral link DigitalOcean Affiliate partner

Also: E2E Networks

Buy this GPU

Own the VRAM

Compare current GPU prices first on our GPU price table and measured speeds.

Amazon Affiliate partner

We may earn from links marked Affiliate partner or Referral link. Disclosure.

New

LLM APIs

Don't want to run it yourself? Compare per-token prices for open models →New

AI agent sandboxes

Isolated sandboxes for agents that run code →

Rent a GPU or buy one

We may earn from some links (disclosure).

Also: E2E Networks · Lambda · Lightning AI · TensorDock · Salad · Hyperbolic · Together AI · B&H Photo · Newegg · Micro Center

Logos are trademarks of their respective owners.

Cheapest way to run it

ModelNeeds (4-bit, 8K)Smallest GPUCheapest rentalAPI $ in / out per 1M
Llama 3.1 8B Instruct6.0 GBRTX 4060$0.06/hr$0.02 / $0.05
DeepSeek-R1-0528-Qwen3-8B6.2 GBRTX 4060$0.06/hr$0.06 / $0.09
Gemma 4 12B7.8 GBRTX 3060$0.06/hr–

See every model: own GPU vs rent vs API →

Weights + KV cache from real Hugging Face file sizes, 90% fit rule, every number sourced. How we calculate →

Join the fit-alerts waitlist

Alerts are not running yet. Join the waitlist and pick your card once; when alerts launch we will email you once to confirm, and only after you confirm will you get the first alerts. Unsubscribe any time.

We store only your email and the GPU or VRAM you pick (plus the page you signed up from). A one-way hash of your IP is used for rate limiting only, for about 15 minutes. Unconfirmed sign-ups are deleted after about 30 days. Never sold or shared.