colibri

Run large AI models across your computers and servers. Powered by Colibri. Pure C.

PHP Framework

My GLM-5.2 work on DGX Spark (GB10) — built with colibr CACHE_ROUTE PR #199, measurements, scripts.

Head Mouse made in Arduino for disability-aid or fun purposes.

Arianna

Reset your Colibri License in a WordPress Website. Impossible to do in the UI without this plugin. This was supplied to me by Colibri's support.

Alternative VK micro-messenger/music-player for Windows 10 Mobile. RU zone preferred.

Running GLM-5.2 (744B MoE) on an NVIDIA GB10 with colibri — conversion stats, benchmark data, and a real GPU-detection bug found and fixed along the way

Faster expert loading for Mixture-of-Experts LLM inference on small machines: parallel miss loading and bit-exact expert-major prefill for colibri (Qwen3-Coder-30B). TTFT 376 s -> 30 s.

An operator-friendly toolkit for deploying Colibri as a persistent CPU/NVMe-backed, OpenAI-compatible AI service while keeping the GPU available for other workloads.

Distributed blade & backplane architecture for frontier-scale MoE models. 4–6x RK3588 ARM nodes, 12–18 NVMe channels, ~16–24 GB/s storage BW.

Colibri framework Documentation

Dashboard for on-train data visualization in 3D - working PoC for Deutsche Bahn Hackathon 2017

Repository of the Colibri programming language

Control board for the Necta Colibri vending machine.

my first project

Running local models

Experimental low-memory LLM inference: AirLLM readers, Transformers forward and native Colibri cache policy. Active WIP; no production or speed guarantees.

Lightweight C inference engine for large mixture-of-experts models, streaming routed experts from NVMe and caching them across RAM and VRAM. Includes adaptive expert placement, CPU and optional GPU backends, KV reuse, a CLI, OpenAI-compatible serving, and a web dashboard.