# JustVugg/colibri

Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦

Repository: https://github.com/JustVugg/colibri
Canonical: https://ross.abutalabs.com/products/colibri
Homepage: https://justvugg.github.io/colibri
Language: C
License: Apache-2.0
License Family: permissive
Last push: 2026-08-26T21:25:10+00:00

## Health v2 (maintenance only)
Score: 80/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 99, release rhythm 99, longevity 4
- inputs: {"age_days": 63, "days_push": 7, "days_rel": 9, "gap_med": 2, "n_releases_24m": 12}
- flags: young
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 26251, forks 2872 (observed 2026-08-28T04:11:38.877179+00:00)

## What it is
Colibrì is a pure-C, zero-dependency inference engine that runs frontier Mixture-of-Experts models (744B to 2.8T parameters) on consumer hardware by streaming expert weights from disk and treating VRAM, RAM, and storage as a single memory hierarchy. It ships per-model engines with a unified front end (chat, serve, web dashboard) and validates outputs token-exact against reference implementations.

## Use cases
- run trillion-parameter MoE models on a consumer PC
- stream LLM expert weights from NVMe instead of loading into RAM
- self-host large language models without a datacenter
- serve a 744B model from a C engine with no dependencies
- run frontier models on Apple Silicon via Metal
- inspect which MoE experts fire with a live dashboard
- run LLM inference on CPU with limited VRAM

## When to choose
- you want to run very large MoE models on hardware with limited VRAM/RAM
- you need a dependency-free, pure-C inference engine you can audit and hack on
- you want token-exact outputs matching reference implementations
- you're researching inference-side systems: memory tiering, prefetching, CPU/GPU overlap

## When to avoid
- you need guaranteed latency or an SLA for production serving
- you need broad model support beyond the six supported families
- you want a mature ecosystem with extensive tooling and integrations
- you need fast throughput on a single consumer machine — streaming from disk trades speed for capacity

## Facets
- artifact type: library
- maturity: active
- function: llm-inference, llm-training, gpu-computing, cli, http-server, chat-interface
- domain: large-language-models, machine-learning, artificial-intelligence, developer-tools, self-hosted
- platform: windows, cpp, cli, self-hosted, cross-platform
- tags: mixture-of-experts, inference-engine, expert-streaming, memory-tiering, pure-c, zero-dependencies, quantization, int4, nvme-streaming, cpu-inference, apple-silicon, metal, linux, macos, gpu

## Member repositories
- JustVugg/colibri (main) score 80

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:11:38.877179+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-29T16:55:54.458997+00:00, confidence not recorded.
  - readme: https://github.com/JustVugg/colibri (fetched 2026-08-28T04:11:38.877179+00:00, sha 28cfb0c42536)
  - homepage: https://justvugg.github.io/colibri (fetched 2026-08-29T07:52:18.665014+00:00, sha 1305a77ec2f3)
- Data as of 2026-08-30T08:39:29.467469+00:00.