Guide · Memory budgets
Which LLMs fit in 16, 24 or 32 GB?
Counted against measured GGUF sizes, not param-count folklore: 75 of the 149 open models tracked here fit a 16 GB machine at Q4_K_M, 111 fit 24 GB and 119 fit 32 GB. What changes between the tiers, and the shortlist worth running at each one.
At Q4_K_M, the roomiest 16 GB machine tops out around 21B parameters, 24 GB around 36B and 32 GB around 49B. A 70B model (42.52 GB on disk) fits none of the three. The usable pool, not the sticker gigabytes, decides every one of those lines.
How these counts work
A model fits when its measured Q4_K_M file plus a 4K-context KV cache and runtime overhead stays inside the machine's usable memory. Usable is the honest number: a GPU keeps roughly all but 1 GB of VRAM for weights, a Mac exposes about two thirds of its unified memory to the GPU working set, and a CPU-only laptop shares its RAM with the OS and apps. The full math is on the methodology page, the sizes come from real GGUF files with sources, and the VRAM calculator runs the same check interactively for any model and machine.
What fits in 16 GB
- Apple unified memory (Apple M5 (16GB)) ~10.5 GB usable · 68 models fit
- GPU VRAM (Nvidia GeForce RTX 4080 (16GB)) ~15 GB usable · 75 models fit
- System RAM (CPU only) (16GB RAM Laptop (CPU/iGPU only)) ~12 GB usable · 72 models fit
On the roomiest 16 GB reading (Nvidia GeForce RTX 4080 (16GB)), the largest fit is gpt-oss 20B at 11.28 GB on disk, and 66 of the 75 fitting models also run at near-lossless Q8_0. Picks worth starting with:
- Gemma 3 12B 7.3 GB
- Gemma 3 4B 2.49 GB
- Gemma 2 9B 5.76 GB
- Phi-4 14B 9.05 GB
- Llama 3.1 8B 4.92 GB
Full tier page with every interpretation: best LLM for 16 GB .
What fits in 24 GB
- Apple unified memory (Apple M4 Pro (24GB)) ~16 GB usable · 80 models fit
- GPU VRAM (Nvidia GeForce RTX 4090 (24GB)) ~23 GB usable · 111 models fit
On the roomiest 24 GB reading (Nvidia GeForce RTX 4090 (24GB)), the largest fit is Seed-OSS 36B Instruct at 20.27 GB on disk, and 74 of the 111 fitting models also run at near-lossless Q8_0. Picks worth starting with:
- Qwen3 30B-A3B 18.6 GB
- Gemma 3 27B 16.55 GB
- Mistral Small 3 24B 14.33 GB
- Qwen3 32B 19.8 GB
- Gemma 3 12B 7.3 GB
Full tier page with every interpretation: best LLM for 24 GB .
What fits in 32 GB
- Apple unified memory (Apple M5 (32GB)) ~21 GB usable · 99 models fit
- GPU VRAM (Nvidia GeForce RTX 5090 (32GB)) ~31 GB usable · 119 models fit
- System RAM (CPU only) (32GB RAM Laptop (CPU/iGPU only)) ~28 GB usable · 117 models fit
On the roomiest 32 GB reading (Nvidia GeForce RTX 5090 (32GB)), the largest fit is Llama-3.3-Nemotron-Super-49B-v1 at 28.14 GB on disk, and 87 of the 119 fitting models also run at near-lossless Q8_0. Picks worth starting with:
- Qwen3 30B-A3B 18.6 GB
- Gemma 3 27B 16.55 GB
- Mistral Small 3 24B 14.33 GB
- Qwen3 32B 19.8 GB
- Gemma 3 12B 7.3 GB
Full tier page with every interpretation: best LLM for 32 GB .
FAQ
What LLM can I run with 16GB of VRAM or RAM?
On the roomiest 16GB reading (Nvidia GeForce RTX 4080 (16GB), ~15 GB usable), 75 of the 149 open models tracked here fit at Q4_K_M, up to about 21B parameters (gpt-oss 20B, 11.28 GB on disk). A 16GB Mac or CPU laptop exposes less usable memory, so its count is lower; the tier table above shows each reading.
Which LLMs and quants are possible with 32GB?
At 32GB, 119 tracked models fit at Q4_K_M (up to ~49B), and 87 of them also fit at near-lossless Q8_0. 30B-class MoE models are the sweet spot: 30B-class quality with 3B-class speed from the active-parameter count.
Is 32GB of RAM enough for a 70B model?
No. A 70B model needs about 42.52 GB on disk even at Q4_K_M, before the KV cache. No 32GB machine holds that; on the tracked devices, 70B first fits on the 64GB Macs or on multi-GPU rigs.
Do RAM, VRAM and unified memory gigabytes count the same?
No. At the same 16GB sticker, the tracked devices expose different pools for weights: the GPU keeps the most usable, the Mac exposes about two thirds of unified memory to the GPU working set, and a CPU-only laptop shares its RAM with the OS and apps. The counts above use each device's sourced usable figure, not the sticker.