VRAM calculator by precision

Type a model name (pick from the list) or a parameter count such as 8B, 70B, 1.5T or 7e9. The table shows how much memory the weights alone need at each precision.

How we count

• Models we track: FP16 is the BF16 file size, INT8 the Q8_0 GGUF file and INT4 the Q4_K_M GGUF file, all read from Hugging Face. When a model has no plain Q8_0 or Q4_K_M build, the table says so and the note names the file we used. The same data is in our comparison table and as JSON.
• Any other parameter count is an estimate: FP16 = 2 bytes per parameter, FP8 = 1, INT8 (Q8_0) = 1.0625, and INT4 (Q4_K_M) = 0.63 bytes per parameter. That last figure is the average across the 22 models we track that have a real Q4_K_M file; the real range is 0.57 to 0.86, so a specific model can land well away from the estimate.
• FP8 is always an estimate (1 byte per parameter), because most models don't publish an FP8 file.
• Weights are not the whole story. Add the KV cache for your context length and leave about 10% headroom for the runtime. The homepage calculator does the full fit check against your GPU, what can I run works backwards from your card, and the GPU VRAM lookup lists every card at once.

1 GB = 1,000,000,000 bytes. For rental options when a model is bigger than your card, see GPU rental prices and Vast.ai vs Runpod per hour.