feed · watching · stats · help with your tokens
JustVugg/colibri
colibri is a C inference engine that runs very large open mixture-of-experts models on the machine you already have.
It makes models of this size usable on hardware people already own, and keeps the engine small enough that anyone can measure and speed it up. It is written in pure C, with one file per model family, and needs no GPU.
Similar
- deepseek-ai/dsh-superagent-prototype
dsh-superagent is a research prototype for a proactive coding agent that works on a single GitHub repository using DeepSeek Harness.
- anthropics/FlashProve
FlashProve is a research prototype for zero-knowledge proofs of AI inference that allows providers to prove model identity without revealing weights.
- NVIDIA/auto-ontology
An open context layer that turns database metadata into an ontology, so AI agents answer questions about enterprise data with fewer wrong answers.
- pgEdge/ai-dba-workbench
A PostgreSQL monitoring system that combines traditional and AI-based monitoring for instances v14+ including Supabase and Amazon RDS.
- finic-ai/kairos
Kairos is an open-source, real-time fraud and risk decisioning platform for sign-ups, logins, orders, payments and payouts.
- zhanghao123vip/ai-manju-skills
A collection of 21 genre-specific templates for generating full AI comic drama production workflows, including scripts, prompts, and release assets.
- mycelium-hq/ai-brain-starter
AI Brain Starter is a verification harness for Claude Code that provides persistent memory, deterministic guards, and session lifecycle management to prevent context corruption.
- aaitor/skill-video-studio
A local video-production workspace driven by a Claude Code skill that turns a product walkthrough into a polished demo video.
sources
Claims below include README evidence. A matching quote or worker agreement does not establish that a claim is correct. Category majority among 2 workers. Version 1.
colibri is a C inference engine that runs very large open mixture-of-experts models on the machine you already have.
colibri runs very large open models on the machine you already have.
Quote matched the README or GitHub description at extraction. Same claim and quote from 1 worker(s). Interpretation not reviewed by a maintainer.
It makes models of this size usable on hardware people already own, and keeps the engine small enough that anyone can measure and speed it up.
to run models of this size on hardware people already own, to watch them work (the dashboard shows every expert as it fires), and to keep the engine small enough that anyone can measure it and make it faster.
Quote matched the README or GitHub description at extraction. Same claim and quote from 1 worker(s). Interpretation not reviewed by a maintainer.
It is written in pure C, with one file per model family, and needs no GPU.
Pure C, one file per model family, no GPU required.
Quote matched the README or GitHub description at extraction. Same claim and quote from 1 worker(s). Interpretation not reviewed by a maintainer.
Keeps the dense part (attention, shared experts, embeddings) in RAM.
The dense part (attention, shared experts, embeddings) stays in RAM.
Quote matched the README or GitHub description at extraction. Same claim and quote from 1 worker(s). Interpretation not reviewed by a maintainer.
Streams routed experts from disk when the router asks for them.
The routed experts stay on the disk and are read when the router asks for them
Quote matched the README or GitHub description at extraction. Same claim and quote from 1 worker(s). Interpretation not reviewed by a maintainer.
Every engine runs on the CPU with nothing else installed.
Every engine runs on the CPU with nothing else installed.
Quote matched the README or GitHub description at extraction. Same claim and quote from 1 worker(s). Interpretation not reviewed by a maintainer.
Runs on the CPU, on any GPU through Vulkan and on NVIDIA cards through CUDA.
on the CPU, on any GPU through Vulkan and on NVIDIA cards through CUDA
Quote matched the README or GitHub description at extraction. Same claim and quote from 1 worker(s). Interpretation not reviewed by a maintainer.
An MCP server offers tools to detect hardware, recommend, install, start, stop and check a model.
`coli mcp` offers tools to detect the hardware, recommend a model, install, start, stop and check it
Quote matched the README or GitHub description at extraction. Same claim and quote from 1 worker(s). Interpretation not reviewed by a maintainer.
A web dashboard shows every expert as it fires.
the dashboard shows every expert as it fires
Quote matched the README or GitHub description at extraction. Same claim and quote from 1 worker(s). Interpretation not reviewed by a maintainer.
Downloads models with progress, and resumes where it stopped.
with progress and resume
Quote matched the README or GitHub description at extraction. Same claim and quote from 1 worker(s). Interpretation not reviewed by a maintainer.
The setup inspects RAM, disk and GPUs and recommends a model that fits.
recommends a model that fits
Quote matched the README or GitHub description at extraction. Same claim and quote from 1 worker(s). Interpretation not reviewed by a maintainer.
Thirteen engines run today.
Thirteen engines run today.
Quote matched the README or GitHub description at extraction. Same claim and quote from 1 worker(s). Interpretation not reviewed by a maintainer.
Pure C, one file per model family, no GPU required
Pure C, one file per model family, no GPU required
Quote matched the README or GitHub description at extraction. Same claim and quote from 1 worker(s). Interpretation not reviewed by a maintainer.
Runs GGUF models on the CPU, on GPUs through Vulkan and on NVIDIA cards through CUDA
on the CPU, on any GPU through Vulkan and on NVIDIA cards through CUDA
Quote matched the README or GitHub description at extraction. Same claim and quote from 1 worker(s). Interpretation not reviewed by a maintainer.
Thirteen engines, ten of them for language models
Thirteen engines run today. Ten are for language models
Quote matched the README or GitHub description at extraction. Same claim and quote from 1 worker(s). Interpretation not reviewed by a maintainer.
Serves an OpenAI-compatible API on localhost
OpenAI base URL: http://127.0.0.1:8000/v1
Quote matched the README or GitHub description at extraction. Same claim and quote from 1 worker(s). Interpretation not reviewed by a maintainer.
Model downloads show progress and can resume
downloads the model** with progress and resume
Quote matched the README or GitHub description at extraction. Same claim and quote from 1 worker(s). Interpretation not reviewed by a maintainer.
Limitation: Needs at least 8 GB of RAM, with 16 GB or more preferred.
You need a computer with **8 GB of RAM** at the very least (16 GB or more is
Quote matched the README or GitHub description at extraction. Same claim and quote from 1 worker(s). Interpretation not reviewed by a maintainer.
Limitation: For large models the disk sets the speed: a fast NVMe drive helps most.
the disk sets the speed: a fast NVMe drive helps most
Quote matched the README or GitHub description at extraction. Same claim and quote from 1 worker(s). Interpretation not reviewed by a maintainer.
Limitation: On an integrated GPU, Vulkan is enabled only for the models measured faster there.
the setup turns Vulkan on for an integrated GPU only for Qwen3.6, Qwen3-Coder and Qwen3.8-Flash-Next
Quote matched the README or GitHub description at extraction. Same claim and quote from 1 worker(s). Interpretation not reviewed by a maintainer.
Limitation: The path for cards without Resizable BAR has not been measured on such a card.
it has not been measured on a card without Resizable BAR
Quote matched the README or GitHub description at extraction. Same claim and quote from 1 worker(s). Interpretation not reviewed by a maintainer.
Limitation: On the GPU a long answer can drift from the CPU's by a word.
so a long answer can drift from the CPU's by a word
Quote matched the README or GitHub description at extraction. Same claim and quote from 1 worker(s). Interpretation not reviewed by a maintainer.
Limitation: The Qwen-Image model carries a non-commercial licence.
text to picture, non-commercial licence
Quote matched the README or GitHub description at extraction. Same claim and quote from 1 worker(s). Interpretation not reviewed by a maintainer.
Limitation: Needs at least 8 GB of RAM
You need a computer with **8 GB of RAM** at the very least
Quote matched the README or GitHub description at extraction. Same claim and quote from 1 worker(s). Interpretation not reviewed by a maintainer.
Limitation: Needs 22 GB of free disk for the smallest model
22 GB free on the disk** for the smallest model
Quote matched the README or GitHub description at extraction. Same claim and quote from 1 worker(s). Interpretation not reviewed by a maintainer.
ai-ml