LLM VRAM & Hardware Fit Calculator

Check if your GPU or Mac can run local LLMs — VRAM estimates by model size, quantization and context length.

Check Your Hardware

Select your hardware, quantization, and context length. Results update instantly — everything runs locally in your browser.

Hardware type

GPU model

Number of GPUs

Quantization

Context length

Available memory

16.0 GB

Memory bandwidth

288 GB/s

Largest model @ 85% pool

≈ 23B · Q4_K_M

Model Compatibility

25 of 30 models can run on this hardware

Model	Memory	Verdict	Est. speed (tok/s)	Pool used
DeepSeek Coder V2 LiteMoEMid-size · Coding Runs on GPU	9.41 GB	Perfect fit	132	59%
Gemma 3 12BMid-size · General Runs on GPU	7.25 GB	Perfect fit	26.4	45%
Mistral Nemo 12BMid-size · General Runs on GPU	7.36 GB	Perfect fit	26.0	46%
Phi-4 14BMid-size · Reasoning Runs on GPU	8.33 GB	Perfect fit	22.6	52%
Qwen3 14BMid-size · General Runs on GPU	8.76 GB	Perfect fit	21.4	55%
Mistral 7B v0.3Light · General Runs on GPU	4.62 GB	Perfect fit	44.0	29%
Qwen2.5 7BLight · General Runs on GPU	4.84 GB	Perfect fit	41.7	30%
Qwen2.5 Coder 7BLight · Coding Runs on GPU	4.84 GB	Perfect fit	41.7	30%
DeepSeek-R1 Distill 7BLight · Reasoning Runs on GPU	4.84 GB	Perfect fit	41.7	30%
Llama 3.1 8BLight · General Runs on GPU	5.06 GB	Perfect fit	39.6	32%
Ministral 8BLight · General Runs on GPU	5.06 GB	Perfect fit	39.6	32%
Qwen3 8BLight · General Runs on GPU	5.17 GB	Perfect fit	38.6	32%
GLM-4 9BLight · General Runs on GPU	5.61 GB	Perfect fit	35.2	35%
Gemma 2 9BLight · General Runs on GPU	5.72 GB	Perfect fit	34.4	36%
Qwen3 1.7BTiny · General Runs on GPU	1.54 GB	Perfect fit	186	10%
Llama 3.2 3BTiny · General Runs on GPU	2.40 GB	Perfect fit	99.0	15%
Phi-4-miniTiny · General Runs on GPU	2.73 GB	Perfect fit	83.4	17%
Gemma 3 4BTiny · General Runs on GPU	2.85 GB	Perfect fit	79.2	18%
Qwen3 4BTiny · General Runs on GPU	2.85 GB	Perfect fit	79.2	18%
Qwen3 30B A3BMoEPro · General MoE expert offload	17.2 GB	Runs well	76.8	49%
Mixtral 8x7BMoEPro · General MoE expert offload	25.8 GB	Runs well	19.6	73%
Qwen3 32BPro · General Partial CPU+GPU offload	18.4 GB	Runs well	2.90	52%
DeepSeek-R1 Distill 32BPro · Reasoning Partial CPU+GPU offload	18.4 GB	Runs well	2.90	52%
Gemma 3 27BPro · General Runs on GPU	15.3 GB	Tight fit	11.7	96%
Mistral Small 3.1 24BMid-size · General Runs on GPU	13.7 GB	Tight fit	13.2	86%
Llama 3.3 70BFlagship · General Runs on GPU	38.4 GB	Won't fit	—	240%
Qwen2.5 72BFlagship · General Runs on GPU	39.5 GB	Won't fit	—	247%
Llama 4 ScoutMoEFlagship · General Runs on GPU	58.5 GB	Won't fit	—	366%
Qwen3 235B A22BMoEFlagship · General Runs on GPU	124 GB	Won't fit	—	775%
DeepSeek-R1MoEFlagship · Reasoning Runs on GPU	348 GB	Won't fit	—	2178%

All figures are formula-based estimates anchored on the open-source llmfit project. Real throughput varies by runtime (Ollama, llama.cpp, LM Studio, MLX), prompt length, and kernel quality. For measured numbers on your machine, run llmfit bench.

Approximate VRAM by model size

Rough planning estimates for ~8k context, Q4 weights plus KV cache. Real usage varies by runtime, context length and MoE offload.

Model class	Quant	Approx. VRAM	Typical hardware
7–8B dense	Q4_K_M	5–6 GB	8 GB GPU or Mac 8–16 GB
7–8B dense	Q8_0	9–10 GB	12–16 GB GPU
13–14B dense	Q4_K_M	9–11 GB	12–16 GB GPU
27–32B dense	Q4_K_M	16–20 GB	24 GB GPU (e.g. RTX 4090)
70–72B dense	Q4_K_M	35–42 GB	Multi-GPU or Mac 48–64 GB+
MoE (~8×7B active)	Q4_K_M	20–28 GB*	24–32 GB+; experts may offload

LLM hardware FAQ

How much VRAM do I need to run a 7B LLM?

Most 7–8B models at Q4 quantization need about 5–6 GB for weights plus KV cache at a short context. An 8 GB GPU can often run them; longer contexts or Q8 need closer to 10 GB.

Can I run a 70B model on a single GPU?

Usually not with full weights on one consumer card. Q4 70B-class models often need ~35–42 GB. You typically need multi-GPU, heavy offload to system RAM, or a Mac with large unified memory.

What quantization is best for a local LLM?

Q4_K_M is the common default: good quality-to-size trade-off. Use Q5/Q6 when you have spare VRAM; Q8 when quality matters more than speed or when the model is small enough to fit.

Does Mac unified memory work for LLM inference?

Yes. Apple Silicon can run many models because CPU and GPU share memory. A 32–64 GB Mac can often host mid-size models that would need multi-GPU on discrete setups, with bandwidth-limited speed.

Does longer context length use more VRAM?

Yes. The KV cache grows with context length and batch size. A model that “fits” at 4k tokens may OOM at 32k — always leave headroom beyond weights alone.