chat completion
This list is the conversation. Generate appends an assistant turn and streams into it; Stop keeps whatever has already arrived and the turn stays editable.
raw completion
No chat template: this text is the prompt, tokenized as typed. Completion models usually expect a trailing space.
chat template
Passed to the Jinja template as extra variables, e.g.
enable_thinking for MiniCPM / Qwen-style templates.
Messages with images render one media marker per
image (<__media__>), in attachment order; that marker is what the projector
replaces with image tokens at generation time, so the preview shows the marker itself.
recommended models
Driven by web/models/recommended.json — a plain static file, so this list can
change without rebuilding the page. Downloads take the same route as the Models tab (4 MiB chunks
into IndexedDB, resumable, with the publisher's sha256 checked on the way in).
loading…
model library
from a URL
Downloaded in 4 MiB chunks into IndexedDB: pausable, resumable, deletable.
from this computer
Left unchecked the file is read in 4 MiB slices at load time and never duplicated — you only re-pick it after a reload. Checked, a copy is written to IndexedDB (one copy in total) so the model survives a reload on its own.
- no models yet — add one above