Colibri, a tiny c engine to run frontier mixture of experts models on a normal machine, topped…
colibri, a tiny c engine to run frontier mixture of experts models on a normal machine, topped github trending on september 14 with no cloud and no api. the largest models are being shrink-wrapped into single-file runtimes.
local moe inference just became a hobbyist project.
Context
The Colibri README, read on 4 October 2026, describes a pure C engine with zero engine dependencies that streams experts from disk, with one C file per model family and seven families running: GLM-5.2 at 744B, GLM-5.3-Flash, Inkling, Kimi K3, DeepSeek V4 Flash, Qwen3.6 and OLMoE. It lists the Apache 2.0 license. Its author-reported speeds include 4 tokens a second for the 744B model on a stated configuration, 5.8 to 6.8 on six RTX 5090s, about 1.8 on a 128 GB CPU-only desktop, 1.07 on a single RTX 5070 Ti and 0.05 to 0.1 on a 25 GB box, and it states no SLA on speed. Gigazine of 10 July 2026 reports it running GLM-5.2 on a 25 GB PC.
The speeds are the author's, tied to specific configurations, and not comparable across rows. Normal machine holds for some models at very low speeds, and frontier here means large open-weight mixture-of-experts models, not a hosted-model comparison. It is one C file per family, not one file overall. The project existed by 10 July 2026, and no dated evidence for topping GitHub trending on 14 September was found, so that is unverified. Local mixture-of-experts inference becoming a hobbyist project is the author's opinion.
Watch next
- A dated GitHub Trending capture and independent tokens-per-second tests.
Sources
- Colibri README (GitHub)github.com
- Colibri runs GLM-5.2 (Gigazine, 10 Jul 2026)gigazine.net
Provenance
The note above is reproduced unedited from the original post, first published on Threads on 19 September 2026 at 04:19 IST. Sources are the papers and datasets the note draws on.
View the original post ↗Embed this note
More notes
The air is now being asked to keep its own ledger
the air is now being asked to keep its own ledger: ecmwf’s aifs compo becomes the first ai model to forecast atmospheric composition globally every three hours, cleanair simulates 365 days of pm2.5 over china in ten seconds, and a unified framework maps six pollutants at one kilometer across the whole country. the air now files its own composition report.
read the note →The current is now being asked to draw its own map
the current is now being asked to draw its own map: china’s langya 2.0 predicts six ocean phenomena including internal waves and mesoscale eddies, a deep net called wenhai resolves eddies globally with air sea flux formulas built in, and scripps infers surface currents from the way temperature patterns deform in satellite images. the ocean now files its own circulation report.
read the note →The soil is now being asked to report its own carbon
the soil is now being asked to report its own carbon: a nix color sensor paired with generative data augmentation predicts soil organic carbon without a lab, random forest drives 74 percent of soil health mapping studies, and sentinel 2 tracks five year carbon change across france and italy from 922 samples. the dirt now files its own carbon account.
read the note →