colibri
Runs huge mixture-of-experts AI models on an ordinary PC by reading parts of the model from disk; pure C, CPU-first, with a chat dashboard and local API.
colibri (v2.1.0, 31 Oct 2026, ~47k stars) is a pure-C engine that runs large open models such as Qwen3.6, DeepSeek V4 Flash and GLM-5.2 on your own machine, keeping the active part in RAM and streaming the rest from disk. It gives a terminal chat, a browser dashboard and OpenAI- and Anthropic-compatible local APIs. Apache 2.0; each model keeps its own licence. Needs at least 8 GB of RAM and 22 GB of disk for the smallest model. Small models reach about 6-14 tokens a second on a CPU; the giant ones often run at 1 token a second or slower without big hardware.
The facts
| Question | Answer | Source | Checked |
|---|---|---|---|
| Runs on my machine? | Yes. Yes; models run on your own computer, with an optional local server for other apps | github.com | Checked today |
| Code licence | Open source. Apache 2.0 | github.com | Checked today |
| Model weights licence | With conditions. Depends on the model. The README says GLM-5.2 is MIT and Qwen-Image-2.1 is non-commercial only. On Hugging Face the default small model Qwen3.6-35B-A3B is Apache 2.0, DeepSeek V4 Flash and MiMo-V2.6 Flash are MIT, and Kimi K3 and Qwen3.8-Flash-Next use custom licences; Muse Spark Open uses Meta's licence agreement | github.com | Checked today |
| Commercial use | With conditions. The engine is Apache 2.0, so yes; most chat models allow it, but Qwen-Image-2.1 does not and some models have custom terms to check | github.com | Checked today |
| Paid services needed? | None required; the engine and the models it downloads are free | github.com | Checked today |
| Hardware | At least 8 GB of RAM (16 GB or more is better) and 22 GB free disk for the smallest model; no GPU needed. Big models need 16-128 GB of RAM and a fast SSD | github.com | Checked today |
| Operating systems | Windows, macOS and Linux, with a one-step setup and prebuilt releases for each | github.com | Checked today |
| Setup effort | Medium; a setup script picks and downloads a model for you (double-click on Windows, a few terminal commands on macOS and Linux), but downloads are large | github.com | Checked today |
Each fact has its own date. Unknown means nobody has published it. We read their documentation; we didn’t install it. Running it yourself costs your hardware, electricity and setup time. Report a correction.
No licence changes since we started watching. Licence changes: follow colibri and we email you if its terms change.
What it replaces, job by job
| Paid tool | Jobs | How far it goes |
|---|---|---|
| ChatGPT Plus $20/month |
| 1 of 3 jobs. Local chat with open models, private and free once downloaded. The models that fit in ordinary RAM run at about 6 to 14 tokens a second on a CPU; the big ones that stream from disk often manage 1 a second or less unless you have lots of RAM or a GPU. Downloads run from 19 GB to over 1 TB, the big models are heavily compressed, and there is no phone app or cloud sync. Answer quality depends on the model you pick. |