Skip to main content
Tool · Local inference

LLM VRAM Calculator

Choose the memory on each GPU, your card count, and a quantization level. The calculator estimates which open-weight models fit, then shows the minimum memory and card count for each result.

Your hardware

The default is one 24GB GPU at Q4, a common local-inference setup.

GB per card
GPU count

Total available memory: 24GB

Best default for fitting capable models on consumer hardware.

8 models fit

Models for 1× 24GB at 4-bit

Highest-ranked fit: Gemma 4 31B

ModelPublic scoreEstimated minimumCards neededReference deployment
Gemma 4 31B31B dense60.1920.4GB17.7GB weights, plus 15% allowance1× 24GB3.7GB spare across your setup1× NVIDIA RTX 4090 (24GB) · 24GB totalCheck monthly cost →
Qwen3.6-27B27B dense52.7917.7GB15.4GB weights, plus 15% allowance1× 24GB6.4GB spare across your setup1× NVIDIA RTX 4090 (24GB) · 24GB totalCheck monthly cost →
Gemma 4 12B12B dense46.247.9GB6.9GB weights, plus 15% allowance1× 24GB16.1GB spare across your setupNo matching catalog config yet
GPT-OSS 20B20B MoE · 20B active41.9113.2GB11.4GB weights, plus 15% allowance1× 24GB10.8GB spare across your setupNo matching catalog config yet
ZAYA1-8B8B dense33.45.3GB4.6GB weights, plus 15% allowance1× 24GB18.7GB spare across your setupNo matching catalog config yet
Phi-4-Reasoning 15B15B dense9.9GB8.6GB weights, plus 15% allowance1× 24GB14.1GB spare across your setup1× NVIDIA RTX 4090 (24GB) · 24GB totalCheck monthly cost →
Mellum2-12B-A2.5B-Thinking12B MoE · 2.5B active7.9GB6.9GB weights, plus 15% allowance1× 24GB16.1GB spare across your setupNo matching catalog config yet
Ornith-1.0-9B9B dense5.9GB5.2GB weights, plus 15% allowance1× 24GB18.1GB spare across your setupNo matching catalog config yet

What the result does not promise

A memory fit does not guarantee a specific token speed or context length. The estimate covers weights and a fixed runtime allowance; the KV cache grows with context and concurrent requests. MoE rows use total parameters for memory even when far fewer parameters are active per token.

LLM VRAM calculator FAQ

How much VRAM does an LLM need?

An LLM needs enough memory for every model weight, plus runtime overhead and the KV cache. At Q4, a useful starting estimate is about 0.57GB per billion parameters before overhead. This calculator adds 15% for runtime use; longer contexts, larger batches, and some inference engines need more.

Can I run a 70B LLM on 24GB VRAM?

Not as a full 70B model on one 24GB card under this estimate. A 70B model needs about 45.9GB at Q4 after the calculator’s runtime allowance, so two 24GB cards are the practical minimum. CPU or system-memory offloading can work, but it usually reduces generation speed.

Do MoE active parameters reduce VRAM requirements?

Active parameters reduce the work performed for each generated token, but they do not remove the other experts from memory. VRAM sizing therefore uses total parameters, not active parameters. A 120B MoE with 12B active can compute like a smaller model while still storing roughly 120B parameters.

Is Apple unified memory the same as GPU VRAM?

Unified memory can hold model weights much like VRAM, so the same capacity estimate is a useful first check. It is not a performance equivalence: memory bandwidth, inference software, quantization support, and the share reserved by macOS all matter. Leave extra headroom when sizing a Mac for local inference.

Next checks