Models that fit your GPU

The newest Hugging Face text models that fit a single 24 GB card (RTX 4090) once Trapetum presses them to 4-bit. Synced daily, most recent first. This is the same catalog your local Trapetum server fetches and compresses on demand, hover a tag to see what each model does best. See the measured PPL and energy per model, or watch a 4-bit model race its fp16 twin live.

Loading the catalog…

How to get one of these models

Click Get model on any model for the steps: install the Trapetum server, paste the model id into its Models tab, and it downloads the fp16 weights from Hugging Face and presses them to 4-bit on your own GPU, then serves them on localhost:8088. Nothing is precompressed on our side; the compression runs on your machine.

~/.trapetum/<model-id>/
  model.cbk        # the 4-bit weights
  tokenizer.json
  config.json
  meta.json