llama.cpp adds in-app model discovery and download resume
A built-in Hub browser, resumable downloads, and an HTML sanitization fix
The facts
- b11714 adds a Discover Models surface to the llama.cpp web UI for browsing and downloading GGUF models from Hugging Face Hub
- The feature is off by default, gated by an 'Enable Discover Models' setting that also toggles a Hugging Face Hub setting
- Downloads can now resume from disk after being paused, with the router persisting incomplete-download state across restarts
- A sanitization fix strips position:fixed and position:absolute from model card HTML to stop overlay-style UI spoofing, while preserving KaTeX math rendering
Llama.cpp's pre-release build b11714 adds a "Discover Models" surface to its bundled web UI. It lets users browse and download GGUF models from the Hugging Face Hub without leaving the app, according to the release notes on GitHub.
The new view sits inside the existing models manager dialog. It offers a curated catalogue, search, and a details pane showing readme and Hub stats. Downloads now get a progress bar in place of the plain one previously used. The feature is off by default: a new "Enable Discover Models" setting gates both browsing and downloading. Toggling it also flips a companion Hugging Face Hub setting, so the two stay in sync.
The release notes describe substantial plumbing behind the download flow. Paused downloads now resume from disk rather than restarting. That's because the router tracks which tags are mid-download and survives a restart with that state intact. Draft sidecar files, used for speculative decoding, are now clustered under their matching main model in the UI. A main quant download will also automatically fetch a quant-matched draft in the background, so speculative decoding is ready once the main weights land.
The notes also mention a fix for a sanitization gap. A malicious model card's raw HTML could previously draw a fixed-position overlay across the app. The sanitizer now blocks this by stripping position:fixed and position:absolute from inline styles, while leaving KaTeX math rendering intact.
Several changelog entries address naming ambiguity in sidecar files, such as drafts and mmproj (multimodal projector) files. Tokens like "MTP" or "mmproj" can appear as either a whole-word marker or buried inside a longer filename. The release tightens the rules so only whole-word matches on non-GGUF-adjacent names count. That prevents a model like "Phi-2-OmniMatrix" from being mistaken for a sidecar.
The notes credit numerous commits to "Assisted-by: pi" alongside named models such as GLM-5.3, GLM-5.3-Flash and DeepSeek-V4.1-Flash. The notes offer no further explanation of what that attribution denotes.
Why it matters
For teams self-hosting via llama.cpp's UI, model acquisition now looks more like a package manager than a manual wget-and-place workflow. The overlay-spoofing fix matters more than it sounds. Any tool rendering untrusted model card HTML faced the same trick. Others pulling in Hub metadata should check their own sanitizers.
Primary source
llama.cpp releases — github.comWritten from the primary source above by The Neural Net’s automated desk (claude-sonnet-5, theneuralnet/pipeline@1) without human review. How this works. Corrections: check the source first — if we got it wrong, the source link is the record.