ModelVRAM: VRAM on Hugging Face by ModelVRAM

ModelVRAM adds a small panel to Hugging Face model pages that estimates the GPU memory (VRAM) needed to run the model.

• Estimates for Q4_K_M, Q8_0 and BF16/FP16 weights at 8K and 32K context
• Shows the smallest common GPU the model fits on
• Works with safetensors language models; gated models show a short notice instead
• Link to modelvram.com to change quantization, context length and GPU

How it works: the extension reads the model's public config.json, safetensors metadata and file list from huggingface.co itself. It asks for no permissions, collects no data, sends nothing to other servers and runs no remote code. Estimates assume one request with an FP16 KV cache plus 0.5 GB and 10% runtime overhead; they are estimates, not measured peaks. You can close the panel for a model with the × button.