The largest AI models are usually out of reach for ordinary computers — they need far more memory than a home machine has. colibri takes a different route: instead of loading the whole model into memory, it streams the parts it needs from disk, just in time. The project describes the idea as "a JIT for weights".
colibri (JustVugg/colibri) is a pure-C, zero-dependency inference engine written by Vincenzo Fornaro and released under the Apache 2.0 licence. Its first commit dates to 1 July 2026, and version 2.1.0 landed on 31 October 2026. It is a one-person project.
The trick is aimed at mixture-of-experts (MoE) models, where only a small set of "experts" is used for each token. colibri keeps the dense parts of the model in RAM and streams the routed experts from NVMe storage as they are needed. That is how it can run models far larger than the machine's memory would normally allow.
Key facts
| Item | Details |
|---|---|
| Repository | JustVugg/colibri |
| Author | Vincenzo Fornaro |
| What it is | Pure-C, zero-dependency inference engine for huge MoE models |
| Licence | Apache 2.0 |
| First commit | 1 July 2026 |
| Latest version noted | v2.1.0 (31 October 2026) |
| How it works | Dense parts in RAM; routed experts streamed from NVMe ("a JIT for weights") |
| Models supported | GLM-5.2 (744B), Kimi K3 (2.8T), Inkling (975B), DeepSeek V4 Flash, Qwen3.8-Flash-Next |
| GitHub stars | 51,687 / 6,260 forks as of 17 November 2026 (RepoRadar) |
The star count has grown fast: 26,662 in late August, 31,868 on 14 September, 39,389 on 4 October and 47,249 on 4 November, reaching 51,687 stars with 6,260 forks as of 17 November 2026. Star counts measure interest, not production use — they show how many developers are watching a project, not how many are running it in production.
Speeds, as reported by the project itself, show the trade-off clearly: about 0.05 tokens per second on a 25 GB development box, about 1.8 tokens per second on a 128 GB CPU desktop, and 5.8–6.8 tokens per second on a machine with 6× RTX 5090 GPUs. GLM-5.2 needs 372 GB of disk space. These are the project's own figures, not independent measurements.
What this means
colibri is best understood as an experiment in access, not speed. Nobody would choose 0.05 tokens per second for real work — but the point is that a model with hundreds of billions of parameters can run at all on hardware that was never meant to hold it. For researchers, students and hobbyists who want to study or probe very large MoE models without a data centre, that is a genuinely new option.
The honest limits matter too. Disk streaming is slow by nature, the project is one person's work, and the speeds quoted are the project's own. colibri opens a door; it does not remove the cost of walking through it.
A note on verification
The facts in this post come from the verified research record for this item, including the repository details and the date-stamped star counts listed above. Star counts change daily, so the figure quoted here is a snapshot as of 17 November 2026, not a live count. Speed figures are the project's own reported numbers.
Source: JustVugg/colibri repository record; star counts via RepoRadar. Last checked: 21 November 2026.
FAQs
What is colibri?
colibri is a pure-C, zero-dependency inference engine by Vincenzo Fornaro that runs huge mixture-of-experts (MoE) models on consumer hardware by streaming routed experts from NVMe storage while keeping the dense parts in RAM.
Which models does colibri support?
The project lists GLM-5.2 (744B), Kimi K3 (2.8T), Inkling (975B), DeepSeek V4 Flash and Qwen3.8-Flash-Next.
How fast is colibri?
Speeds are the project's own figures: about 0.05 tokens per second on a 25 GB development box, about 1.8 tokens per second on a 128 GB CPU desktop, and 5.8–6.8 tokens per second on 6× RTX 5090 GPUs. GLM-5.2 needs 372 GB of disk space.
How popular is colibri on GitHub?
As of 17 November 2026, the JustVugg/colibri repository had 51,687 GitHub stars and 6,260 forks. Star counts measure interest, not production use.
Comments (0)
Share your question or experience about this syllabus. Your comment will appear after review.
No comments yet — be the first to share your thoughts!