llama.cpp b11714 adds Discover Models: browse and download Hugging Face GGUFs from the llama-server web UI, with drafts and resumable downloads

llama.cpp release b11714, published October 17, 2026, ships a Discover Models view in the llama-server web UI. You can now find and download GGUF models from Hugging Face inside the models manager instead of copying a -hf tag into a terminal. The change comes from PR #29584, merged the same day.

Key facts:

• Off by default: a new “Enable Discover Models” setting turns it on, and turning it on also enables the Hugging Face Hub setting, which stays locked while Discover is active.
• Browsing: a featured catalog, debounced search, size-range filters, per-quant download chips with memory-fit notes, sanitized README rendering and Hub stats.
• Draft sidecars: downloading a main quant automatically starts the quant-matched draft model of the best kind the repo ships, preferring DFlash over MTP, so speculative decoding is ready when the main file lands.
• Resumable downloads: pausing keeps the partial file, and unfinished downloads are remembered across restarts.
• Filtered out: decision models (the decision-model tag or a text-classification pipeline) are hidden from every Hub list, because the UI cannot run them.

How downloads behave now

The PR treats a main quant and its drafts as one unit. Drafts never get their own rows. They show up as sub-lines under the main model in the selector, the Discover list and the manager table. Pausing or resuming the main download cascades to its drafts. One cancel confirmation discards the main file and its drafts, and deleting a main quant removes only its own quant-matched sidecars, so drafts that belong to other quants survive.

Under the hood, the server now writes unfinished downloads to unfinished-downloads.json in the model cache and reports them as still downloading. Tag resolution in common/download.cpp also understands <repo>:<quant>-<sidecar> tags for mtp, eagle3, dflash, dspark, mmproj and imatrix. A tag like Q4_0-mtp can therefore pull just the sidecar file.

The API it sits on

If you script model management rather than click, the same features are exposed through llama-server’s model endpoints, documented in the router-mode section of the server README. Router mode starts when you launch llama-server without a model, and it reads models from the LLAMA_CACHE cache, --models-dir or --models-preset.

• POST /models starts a background download. Its sidecars field accepts mmproj, mtp, dflash, eagle3 and dspark, and defaults to ["mmproj"], the same as -hf. Progress arrives over /models/sse.
• Posting the download to /models/unload pauses it and keeps the partial files. DELETE /models cancels it and drops them.
• DELETE /models?model={name}&sidecar={kind} removes one sidecar and keeps the model, as long as the model is not running.

Should you turn it on?

It is useful if you run llama-server as a shared box and want teammates to pull quants without shell access. One caveat: the setting flips Hub access on for the UI, so leave it off on locked-down or offline machines.

For the previous release’s CUDA decode speedups, see llama.cpp b11713. For running one server for several people, see how to scale llama-server for multiple users, and for the server’s other recent additions, see llama-server built-in tools.