Skip to content

Free tool

Local-LLM GPU Checker

Pick your GPU (or type your VRAM) and see exactly which open LLMs you can run locally — at which quantization, with a fit verdict and a rough speed feel. Built on vendor-verified VRAM and real GGUF sizes.

RTX 4090 (24 GB) · Q4_K_M · 8K (typical chat)

You can comfortably run up to Gemma 3 27B at Q4_K_M.

  • Qwen3 32B32.8B · Qwen
    Tight fit
    Q4_K_M21.3 GB· ~39 tok/s (usable)
  • Qwen2.5 32B32B · Qwen
    Tight fit
    Q4_K_M20.8 GB· ~40 tok/s (fast)
  • DeepSeek-R1 Distill 32B32B · DeepSeek
    Tight fit
    Q4_K_M20.8 GB· ~40 tok/s (fast)
  • Gemma 3 27B27B · Gemma
    Runs well
    Q4_K_M17.5 GB· ~48 tok/s (fast)
  • Mistral Small 3 24B24B · Mistral
    Runs well
    Q4_K_M15.6 GB· ~53 tok/s (fast)
  • Qwen3 14B14.8B · Qwen
    Runs well
    Q4_K_M9.6 GB· ~87 tok/s (fast)
  • Qwen2.5 14B14B · Qwen
    Runs well
    Q4_K_M9.1 GB· ~92 tok/s (fast)
  • Phi-4 14B14B · Phi
    Runs well
    Q4_K_M9.1 GB· ~92 tok/s (fast)
  • DeepSeek-R1 Distill 14B14B · DeepSeek
    Runs well
    Q4_K_M9.1 GB· ~92 tok/s (fast)
  • Gemma 3 12B12B · Gemma
    Runs well
    Q4_K_M7.8 GB· ~107 tok/s (fast)
  • Gemma 2 9B9B · Gemma
    Runs well
    Q4_K_M5.8 GB· ~143 tok/s (fast)
  • Qwen3 8B8.2B · Qwen
    Runs well
    Q4_K_M5.3 GB· ~156 tok/s (fast)
  • Llama 3.1 8B8B · Llama
    Runs well
    Q4_K_M5.2 GB· ~160 tok/s (fast)
  • DeepSeek-R1 Distill 8B8B · DeepSeek
    Runs well
    Q4_K_M5.2 GB· ~160 tok/s (fast)
  • Mistral 7B7.3B · Mistral
    Runs well
    Q4_K_M4.7 GB· ~160 tok/s (fast)
  • Qwen2.5 7B7B · Qwen
    Runs well
    Q4_K_M4.5 GB· ~160 tok/s (fast)
  • Qwen3 4B4B · Qwen
    Runs well
    Q4_K_M2.6 GB· ~160 tok/s (fast)
  • Gemma 3 4B4B · Gemma
    Runs well
    Q4_K_M2.6 GB· ~160 tok/s (fast)
  • Phi-3.5 mini (3.8B)3.8B · Phi
    Runs well
    Q4_K_M2.5 GB· ~160 tok/s (fast)
  • Llama 3.2 3B3.2B · Llama
    Runs well
    Q4_K_M2.1 GB· ~160 tok/s (fast)
  • Gemma 2 2B2.6B · Gemma
    Runs well
    Q4_K_M1.7 GB· ~160 tok/s (fast)
  • Qwen3 1.7B1.7B · Qwen
    Runs well
    Q4_K_M1.1 GB· ~160 tok/s (fast)
  • Qwen2.5 1.5B1.5B · Qwen
    Runs well
    Q4_K_M1 GB· ~160 tok/s (fast)
  • Llama 3.2 1B1.2B · Llama
    Runs well
    Q4_K_M0.8 GB· ~160 tok/s (fast)
  • Gemma 3 1B1B · Gemma
    Runs well
    Q4_K_M0.6 GB· ~160 tok/s (fast)
  • Qwen3 0.6B0.6B · Qwen
    Runs well
    Q4_K_M0.4 GB· ~160 tok/s (fast)
  • Qwen2.5 72B72B · Qwen
    Won't fit
    Q4_K_M46.7 GB
  • Llama 3.3 70B70B · Llama
    Won't fit
    Q4_K_M45.4 GB
  • DeepSeek-R1 Distill 70B70B · DeepSeek
    Won't fit
    Q4_K_M45.4 GB

Need a card for bigger models? See our GPU + mini-PC guides.

Estimates for DENSE models, anchored to real GGUF sizes: ≈ params × bytes-per-weight (Q3 ~0.43 → Q4 ~0.55 → Q8 ~1.06 → FP16 2.0) + a KV-cache factor for your context. Tokens/sec is a rough bandwidth-class “feel” (single user, short context) — not a benchmark; real speed depends on your engine (llama.cpp / Ollama / vLLM / MLX), batch size and context. Apple usable memory ≈ 72% of total. AMD (ROCm/Vulkan) + Intel (IPEX) ecosystems are less mature than CUDA. Re-verify before buying.

Frequently asked

How much VRAM do I need to run a model locally?

Rule of thumb: multiply the model's parameter count (in billions) by the bytes per weight for your quantization — 2 GB/B for FP16, 1 GB/B for 8-bit, ~0.5 GB/B for 4-bit (GGUF Q4_K_M) — then add about 20% for the KV cache, activations and framework overhead. So an 8B model at 4-bit needs roughly 8 × 0.5 × 1.2 ≈ 5 GB, and a 70B model at 4-bit needs about 42–46 GB.

Can I run a 70B model on an RTX 4090 (24 GB)?

Not on a single card at usable quality. Llama 3.3 70B at 4-bit (Q4_K_M) is about 39 GB of weights and ~42–46 GB with context, so it needs roughly 48 GB — typically two 24 GB cards (dual 3090/4090) or an Apple machine with 64 GB+ unified memory. On one 4090, run a 32B model (Qwen2.5 32B or DeepSeek-R1 32B) at 4-bit instead — that fits comfortably.

What's the best LLM for a 16 GB GPU (RTX 4080 / 5080 / 4060 Ti 16GB)?

At 4-bit you comfortably fit 7B–14B models — Llama 3.1 8B, Qwen2.5 14B, Phi-4 14B, Gemma 3 12B — with room for long context. A 24B (Mistral Small 3) is tight but workable at short context; 27B (Gemma) is borderline. Note the 4060 Ti 16GB has a narrow memory bus, so it fits the same models but generates noticeably slower than a 4080/5080.

Does Apple unified memory count as VRAM for LLMs?

Yes, and it's one of Apple Silicon's biggest advantages — the GPU shares the whole unified memory pool. But macOS reserves part of it for the system, so the practical budget for model weights is roughly 70–75% of the total. A 64 GB Mac gives ~48 GB usable (enough for a 70B at 4-bit), and 128 GB gives ~96 GB usable. You can raise the GPU memory cap with a sysctl tweak if you need more.

FP16 vs 8-bit vs 4-bit — which quantization should I use?

For local chat, 4-bit (Q4_K_M) is what almost everyone runs: it cuts VRAM ~75% versus FP16 while losing only ~1.5–2 points on benchmarks. 8-bit is near-lossless (under ~1 point) but doubles the memory of 4-bit. FP16 is reference quality but rarely worth it locally — it's mostly for training/fine-tuning, not inference.

Why does longer context need more VRAM?

Beyond the fixed weight memory, the model stores a KV cache for every token in the conversation, and that grows with context length. A short 4K chat adds only ~12% overhead, but a full 128K-token context can add 50–60% on top of the weights — which can push a model that 'fits' at 8K into 'won't fit' at 128K.

How to check if a model fits your GPU

Whether a specific model runs on a specific card comes down to one question: does the model, at your chosen precision, fit in the card's VRAM with room for the context. This tool checks that fit; here is what it is checking.

Precision changes the answer

The same model can fit or not depending on how it is quantized. At 16-bit it needs about 2 GB per billion parameters; at 4-bit, about 0.5 GB. Dropping precision is the most common way to make a model that does not fit suddenly fit, at a small quality cost.

Context needs headroom

The weights are only the baseline. The KV cache that holds your prompt and the reply grows with context length and can add several gigabytes, so a model that just barely fits its weights may fail on a long prompt. Leave a margin.

If it does not fit

Quantize further, pick a smaller model, shorten the context, or offload some layers to system RAM at the cost of speed. For a full sizing walkthrough, see the GPU for LLM tool.

Related reading

Go deeper on how this works and what to pick.

Newsletter

Liked the tool? Get the signal.

One weekly email on the AI + hardware that actually matters — from the people who build these calculators.

Free · unsubscribe anytime · no spam.