site: a searchable model list, a brio section, and a theme switch - #1635
Merged
Merged
Conversation
added 2 commits
September 20, 2026 15:32
The landing page still described eight families, listed DeepSeek and Qwen3 as planned when both have been running for a while, and showed the models as a three-column card grid whose cards were as tall as their longest description. Brio mode was not mentioned at all. What changed - Models are a list, not a grid: one row per family with the description, the tags, total and active parameters, the container size, the engine file, and the command that fetches the weights with a copy button. A search box and five tag filters narrow it in place. Nine families, up from the eight the page claimed, with DeepSeek V4.1 Flash no longer nested inside the V4 card. - Every figure comes from the family table in the README or, where the README is silent, from the model's own card upstream. The download commands are the ones in each model's doc, and each row links to the doc that exists. - A brio mode section: what it is for with three worked examples, the measured comparison against generating the same answer, the request and the reply of POST /v1/brio, and the calibration figures that are the point of the mode. - A light and dark theme. The mechanism is the one sparshsingh12312 proposed in #1551: the same colibri_theme key, the same data-theme attribute set before first paint so the theme never flashes, the same D shortcut, and the operating system preference followed only until the visitor chooses. - Figtree for the page, Bytesized still for the wordmark, and the favicon is now the mark in its own four colours instead of flat ink. The page stays what it was: one file, no build step, no framework. The filtering, the theme and the copy buttons are about sixty lines of plain JavaScript at the bottom. Note for #1551: that pull request rewrites the same file, so one of the two will have to be rebased on the other. Its dark-mode design is the one kept here.
Two things the first commit got wrong. The contract in c/tests/test_family_registry.py wants site/index.html to print the shipping version and to name every registered family, and the rewrite dropped both. The version and the family names are now read from c/version.py and c/family_registry.py when the page is generated, so they cannot drift again: the footer says "Currently shipping v1.11.0" and the models block names the nine engines the way the docs name them. The hero showed a mocked-up terminal. It is now a recording of a real session, 31 seconds: coli chat attached to a running server, a question answered and stopped with Ctrl-C, then /brio with three options and the model scoring them, then the same thing in the dashboard, ending on the result. Every number on screen came out of the engine, including the ones that are not flattering: the chat arm ran at 0.64 tok/s because the server was reading experts from disk under load, and the brio call took 31 seconds for 4 tokens read and 0 generated. 440 KB of VP8 for 31 seconds, muted and looping, with a poster frame so there is something to look at if it does not autoplay.
This was referenced Sep 20, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What this is
The landing page had drifted. It said eight families run today, listed DeepSeek
and Qwen3 as planned when both have been running for a while, nested DeepSeek
V4.1 Flash inside the V4 card, and did not mention brio mode at all. The models
were a three-column card grid, so every card was as tall as the longest
description in its row.
What changed
The models are a list. One row per family: description, tags, total and
active parameters, the container size, the engine file, and the command that
fetches the weights with a copy button next to it. A search box and five tag
filters narrow the list in place. Nine families now, each linking to the doc
that actually exists for it.
The figures are sourced. Every number comes from the family table in the
README or, where the README is silent, from the model's own card upstream. One
discrepancy surfaced while doing this and is worth a separate look: the README
row for GLM-5.3-Flash says 321B / 40B while Z.ai's card says 320B total and 18B
active, and 40B is also GLM-5.2's figure on the row above. The page uses the
upstream number.
A brio mode section. What the mode is for, with three worked examples
(triage, filling a schema, many questions on one document), the measured
comparison against generating the same answer on qwen36, the request and the
reply of
POST /v1/brio, and the calibration figures, which are the actualpoint of the mode. It also says plainly where brio does not help.
A light and dark theme, using the mechanism @sparshsingh12312 proposed in
#1551: the same
colibri_themekey, the samedata-themeattribute set beforefirst paint so the theme never flashes, the same
Dshortcut, and the operatingsystem preference followed only until the visitor chooses for themselves.
Type and mark. Figtree for the page, Bytesized still for the wordmark, and
the favicon is now the mark in its own four colours instead of flat ink.
What did not change
One file, no build step, no framework, no dependency. The filtering, the theme
and the copy buttons are about sixty lines of plain JavaScript at the bottom of
the page. The expert atlas canvas, the hardware numbers and the contributing
section are the same content, restyled.
Checked
Dshortcut switch them, the choicesurvives a reload, and
Dtyped into the search box types adinstead offlipping the theme.
Conflict to resolve
#1551 rewrites the same file, so one of the two has to be rebased on the other.
Its dark-mode design is the one kept here.