Kolibri-1 against the models it says it beats
On October 3, 2026 Aleph Alpha released Kolibri-1, an open-weight German/English model: 78B parameters, 3.46B active per token, up to 1M tokens of context, Apache 2.0. Its model card says it leads the mixture-of-experts models of its size. It has no API anywhere yet, so we gave it two prompts in Tesseracted Labs' free browser chat, then sent the exact same text to four of the models in that comparison through OpenRouter. One request each, ten files, all playable here.
Kolibri Rush
One prompt, five models: a hummingbird flies through a neon city at night, collects nectar and dodges drones.
The five Kolibri Rush builds running at once. Press Autopilot and the same little bot plays every window (it only reads the game state the prompt asked each model to expose).
Written in about 20 seconds. As delivered it froze on the first frame (it calls ctx.circle, which does not exist); we changed that one call to ctx.arc. It runs, but the city skyline is missing and the flowers crawl at 8 pixels per second (per-frame and per-second speeds are mixed up), so it is barely playable. Original kept untouched.
Same size class as Kolibri (3B active). Starts first time: skyline with lit windows, flowers, drones, speed-up. One bug left as it is: with the keyboard the bird never stops climbing, so play it with the mouse. No edits.
As delivered it could not start: Space and click were never wired to the start screen. We added that (two lines, marked in the source). Then it plays, with a lit skyline. Original kept untouched.
Runs out of the box, but the city is rows of grey towers across the whole screen. No edits.
The most polished of the five once it starts. As delivered the canvas stayed at 300x150 (resize() was never called at start); we added that one call. Original kept untouched.
Kolibri-1 · Aleph Alpha, untouched original · Kolibri-1's answers as text, exactly as written · Nemotron 3 Super, untouched original · Qwen3.8 27B (dense), untouched original
Inside a Mixture-of-Experts model
Same prompt to all five: an interactive page that shows how Kolibri's own architecture routes each token to 6 of 384 experts plus a shared one, layer by layer, with a usage heatmap.
Written with extended thinking on in the free demo, which stopped it at its output limit: the file ends halfway through a function, so the canvas stays empty. That is the demo's cap, not necessarily the model. Untouched.
Complete: token chips, the 384-cell expert grid, live stats and a heatmap. No edits.
Shows nothing: the script has a syntax error (Unexpected token 'var'). Untouched.
Draws the token and its connections to the chosen experts, but not the 384-expert grid. No edits.
The best of the five: the token travels through all 50 layers, experts light up, the heatmap fills. No edits.
Kolibri-1's answers as text, exactly as written
The exact prompts are for Ruocco Academy members. Every prompt word for word, with model, settings, time and cost, the full clean source of every demo as a ZIP, and step-by-step notes on how each video was made. The demos on this page are free to use for everyone.
Join the Academy