llmman

· 3 min read

Find and pull models from the web UI

The web UI has a Models page now: popular models from Docker Hub and Hugging Face, a search of both, and a card per model with every tag or quantization you can pull, its size, and whether it fits your machine.

Pulling a model from llmman’s web UI used to mean knowing its full reference already, typed into a box. The UI now has a Models page, where you find a model first and then pull it.

The Models page: popular models from Docker Hub on the left, and the card for unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF on the right, with its quantizations, their sizes, a fit line and a Pull button

What’s on it

It opens on what most people pull: Docker’s own ai/ models on Docker Hub, then the most downloaded GGUF text-generation repos on Hugging Face, the kind llama.cpp serves. Type in the box and it searches both registries instead. Pulled lists what this machine already has.

Selecting a model opens its card:

  • Every tag or quantization pull can take, each with its size. For a GGUF repo on Hugging Face those are its quantizations, and the one a plain llmman pull would pick is marked default.
  • Whether it fits. Each size is weighed against this machine’s model memory: green under 60%, yellow under 90%, red past that. It’s weights only, so a green model can still run out of room at a long context, and the card says so.
  • Pull, with the same progress /api/pull streams anywhere else. Once it’s here, the card lets you chat with it, unload it or delete it.
  • The README, the model card from Hugging Face or the overview from Docker Hub.
bartowski/SmolLM2-135M-Instruct-GGUF pulling from its card, 7.7 MB of 101 MB, with its quantizations and README below

What it shows is what you get

A list of quantizations is only useful if picking one pulls that file. So the card doesn’t guess from file names. Each tag it lists is run back through the same selection llmman pull uses, and kept only if that lands on the file it was listed with: Q4_K would match the Q4_K_M file too, so it isn’t offered. The size includes the vision projector when pull downloads one alongside, and a diffusion or safetensors repo is a single download, because that’s how pull treats them.

Same results as the CLI

The page doesn’t talk to Hugging Face or Docker Hub itself. It asks llmman serve, and /llmman/search is llmman search over HTTP: the same function, the same rows in the same order, so the terminal and the page can’t disagree. The fit check went the other way. It started on the card and then came to the CLI as a FIT column (#603), green, yellow and red at the same 60% and 90%:

$ llmman search qwen3.5 -n 3
NAME                                     PULLS     LIKES    UPDATED         FIT
docker.io/ai/qwen3.5                     616.5K    10       1 month ago     34%
docker.io/ignaciolopezluna020/qwen3.5    121       0        6 months ago    -
hf.co/Qwen/Qwen3.5-9B                    9.0M      2091     7 months ago    29%
hf.co/Qwen/Qwen3.5-4B                    7.8M      996      7 months ago    14%
hf.co/Qwen/Qwen3.5-2B                    4.9M      423      7 months ago    7%

To try it, run llmman serve and open http://127.0.0.1:17434/#/models.

Details are in docs/webui.md and docs/api.md. Questions are welcome at github.com/llmmanorg/llmman.