Can my GPU or Mac run this model?
Updated . By LLM Hardware team · Methodology
Pick a card or a Mac from our hardware catalog and a popular open model. The tool shows whether the weights fit at Q4_K_M, Q8 and FP16 at a short context. For a custom parameter count or a long context, use the VRAM calculator.
NVIDIA GeForce RTX 4090 (24GB): 24 GB usable for Qwen3 32B (32B).
Open in VRAM Calculator for long context, KV-cache type and MoE offload.
| Quant and formula | Needs | Verdict |
|---|---|---|
| Q4_K_M 32 × 0.5 × 1.2 + 2 | 21.2 GB | Fits |
| Q8 32 × 1 × 1.2 + 2 | 40.4 GB | Does not fit |
| FP16 32 × 2 × 1.2 + 2 | 78.8 GB | Does not fit |
Formula: memory (GB) = parameters (B) × bytes per parameter × 1.2 + 2. Bytes: Q4_K_M 0.5, Q8 1, FP16 2. Apple unified memory has 4 GB removed for macOS first. “Tight” means under 10% headroom.
Amazon links are affiliate links. Prices change; check the listing before you buy.
Next steps
- VRAM calculator for custom sizes and 16K–128K context.
- What can I run on my GPU? for model picks by VRAM tier.
- How much VRAM do I need?
- Best GPU for LLMs and Apple Silicon for LLMs.
- Quantization guide for what Q4, Q8 and FP16 trade off.
- All hardware with specs and fit tiers.
FAQ
How does this tool decide whether a model fits?
It uses the same formula as the VRAM calculator: memory (GB) = parameters (B) × bytes per parameter × 1.2 + 2 GB. Bytes per parameter are 0.5 at Q4_K_M, 1 at Q8 and 2 at FP16. The 1.2 factor and the 2 GB reserve cover activations and a short context of about 4K tokens.
Why does a Mac with the same memory fit less than a GPU?
Apple unified memory is shared with macOS. The tool subtracts 4 GB from the listed unified memory before comparing, on top of the formula’s 2 GB reserve. A 24 GB Mac therefore has about 20 GB to work with, while a 24 GB GPU has the full 24 GB.
What does “tight fit” mean?
The estimate fits, but with less than 10% of memory left over. The model loads at a short context, but a long context or a second app on the GPU can push it past the limit. Pick the next smaller quant or a larger card if you need long prompts.
Do MoE models need memory for only their active parameters?
No. All experts stay resident, so the tool counts the total parameters. Qwen3 30B-A3B is counted at 30B, not 3B. Active parameters affect speed, not the memory needed to load the weights.
Does a long context change the answer?
Yes. The verdicts assume a short context. 16K, 32K and 128K contexts add KV cache on top. The VRAM calculator adds an approximate KV allowance for those lengths, so use it when you plan long prompts.
Can I share a result?
Yes. The address bar updates with your hardware and model, for example /can-i-run?hw=nvidia-rtx-4090-24gb&model=qwen3-32b. Copy it or use the share button. Every shared link is the same page with different inputs, not a separate page.