RunsOnMy ModelsGPUs & MacsBy VRAMAPI pricesVRAM calculatorNew

Which AI models fit in my VRAM?

Pick a memory size to see every open model we track that fits at 4-bit (Q4_K_M) with an 8K context, using the same maths as the rest of the site: real GGUF file sizes, KV cache from each model's config, and 95% of the card's VRAM as usable.

8 GB VRAM

133 models fit at 4-bit · 86 still fit at 8-bit

Usable memory: 7.6 GB. Largest that fits: Llama 3.2 11B Vision Instruct (11B, 7.4 GB at 4-bit).

6 cards in our table have 8 GB, including RTX 3060 Ti, RTX 3070, RTX 4060.

12 GB VRAM

154 models fit at 4-bit · 136 still fit at 8-bit

Usable memory: 11.4 GB. Largest that fits: Kimi VL A3B Thinking 2506 (16B, 10.8 GB at 4-bit).

8 cards in our table have 12 GB, including RTX 3060 12GB, RTX 3080 12GB, RTX 4070.

16 GB VRAM

162 models fit at 4-bit · 145 still fit at 8-bit

Usable memory: 15.2 GB. Largest that fits: LFM2.5 24B A2B (24B, 14.3 GB at 4-bit).

14 cards in our table have 16 GB, including RTX 4060 Ti 16GB, RTX 4070 Ti Super, RTX 4080.

24 GB VRAM

195 models fit at 4-bit · 161 still fit at 8-bit

Usable memory: 22.8 GB. Largest that fits: Maxwell V1 Flash (37B, 22.1 GB at 4-bit).

7 cards in our table have 24 GB, including RTX 3090, RTX 3090 Ti, RTX 4090.

32 GB VRAM

201 models fit at 4-bit · 173 still fit at 8-bit

Usable memory: 30.4 GB. Largest that fits: Kimi Linear 48B A3B Instruct (49B, 28.9 GB at 4-bit).

3 cards in our table have 32 GB, including RTX 5090, Mac M4 32GB, Mac M1 Pro 32GB.

48 GB VRAM

205 models fit at 4-bit · 199 still fit at 8-bit

Usable memory: 45.6 GB. Largest that fits: Kimi Dev 72B (73B, 44.6 GB at 4-bit).

4 cards in our table have 48 GB, including RTX A6000, RTX 6000 Ada, L40S.

64 GB VRAM

209 models fit at 4-bit · 204 still fit at 8-bit

Usable memory: 60.8 GB. Largest that fits: gpt oss 120b (117B, 59.5 GB at 4-bit).

3 cards in our table have 64 GB, including Mac M4 Pro 64GB, Mac M1 Max 64GB, Mac M4 Max 64GB.

A model "fits" when total memory is at most 90% of usable VRAM, and is "tight" up to 100%. Apple Silicon uses different default limits; see the Mac guide and the methodology.