site: four model families, not one — and two cards that were plainly wrong

[JustVugg]

The website still said colibrì runs GLM-5.2. It has run four families since v1.3.0, and two model cards contradicted the project outright:

	before	after
Inkling	975B MoE · Planned · "On the roadmap"	975B MoE · Live · "Run it →"
Kimi	K2 · 1T MoE · Planned	K3 · 2.8T MoE · Live · "Run it →"

Telling visitors that Inkling and Kimi are on the roadmap, while the README front page says both run today, is the kind of contradiction someone finds in thirty seconds. The Kimi card also had the wrong model and the wrong size.

The hero was rewritten, not merely widened

The old line worked because it put two incompatible things next to each other — an enormous model, your machine. Replacing that with a range ("744B to 2.8T parameters") informs and stops landing. The contradiction was the message.

These models do not fit in your machine. They run in it anyway.

A Mixture-of-Experts token touches a small fraction of the weights, so colibri keeps that part resident and streams the rest from disk as it is needed. Four families run today — GLM-5.2 (744B), Inkling (975B), Kimi K3 (2.8T) and OLMoE (7B) — in pure C, with zero dependencies. No cluster, no cloud, no API key.

Two reasons for that shape. It matches the voice further down the page — "Weights are not state to hold. They are data to stage." — so the site speaks with one voice instead of two. And the subtitle now explains why it is possible, which was missing entirely: someone reading "2.8T on my PC" assumes marketing, and twelve words about MoE sparsity turn an unbelievable claim into an understandable one.

The numbers do not disappear — they move to the subtitle, where they serve the reader who wants detail rather than the one who is skimming.

Also updated

<title>, <meta description> and both Open Graph tags — what shows in search results and link previews — from "a 744B model" to "frontier MoE models". Footer model list: DeepSeek → Kimi K3.

Deliberately unchanged

DeepSeek and Qwen3 stay "Planned", because #165 and #712 are not merged. The site should not promise what the code does not do.

Text only. No CSS, structure or script changes. The only markup edits are the two cards’ buttons becoming real links now that both models are runnable.