colibri is a pure-C inference engine that runs very large mixture-of-experts open models on ordinary computers by keeping the always-active weights in RAM and streaming experts from disk as the model needs them.
Follow colibri to get an email when its listing changes, such as new pricing or features. One digest a day at most, and you can unsubscribe anytime.
A one-step setup recommends a model for your machine, builds or downloads the engine and serves OpenAI- and Anthropic-compatible endpoints that coding agents can use.
- 01
Expert streaming from disk so hundreds-of-billions-parameter MoE models run in modest RAM
- 02
CPU, Vulkan and CUDA backends with an optional GPU
- 03
Runs published GGUF files as-is, from IQ1_S to Q8_0, F16, BF16 and MXFP4
- 04
OpenAI-compatible /v1 and Anthropic-compatible endpoints for coding agents
- 05
One-step start-here setup with hardware profiling, model recommendation, resumable downloads and a coli ui dashboard
Free
colibri is free and open-source under the Apache-2.0 license with no paid tiers; the project accepts optional donations. You only need your own hardware (8 GB RAM minimum, 16 GB or more recommended, plus disk space for model weights).
Best for
Developers who want to run frontier-scale open MoE models such as GLM, DeepSeek, Kimi or Qwen locally for private coding-agent work without buying datacenter GPUs.
Not ideal for
Latency-sensitive production serving or dense models, since streaming experts from disk trades speed for memory and throughput depends heavily on disk and RAM bandwidth.
An impressively compact way to run giant open MoE models on consumer hardware, with agent-ready APIs that make it a practical local backend for coding agents if you can accept disk-bound speeds.
v2.1.0 (October 31, 2026) added native GGUF support on every engine, Mistral-Large-4, Muse Spark Open, Flint and LFM3-VL, CUDA device chains, speculative decoding via MTP heads, 262k-token contexts for Qwen3.8-Flash-Next and a new chat UI; v2.0.0 shipped October 6, 2026.
As of
Is colibri free?+
Yes - colibri is Free. colibri is free and open-source under the Apache-2.0 license with no paid tiers; the project accepts optional donations. You only need your own hardware (8 GB RAM minimum, 16 GB or more recommended, plus disk space for model weights).
Is colibri open source?+
Yes - colibri is open source. colibri is free and open-source under the Apache-2.0 license with no paid tiers; the project accepts optional donations. You only need your own hardware (8 GB RAM minimum, 16 GB or more recommended, plus disk space for model weights).
Who is colibri best for?+
Developers who want to run frontier-scale open MoE models such as GLM, DeepSeek, Kimi or Qwen locally for private coding-agent work without buying datacenter GPUs.
Who is colibri not ideal for?+
Latency-sensitive production serving or dense models, since streaming experts from disk trades speed for memory and throughput depends heavily on disk and RAM bandwidth.
What are the best colibri alternatives?+
The closest colibri alternatives on ai.dosa.dev are Copperhead, NVIDIA Personal AI Router (PAIR), WebLLM - all listed under Hardware & Edge AI.
Who makes colibri?+
colibri is developed by JustVugg. It is listed in the Hardware & Edge AI category on ai.dosa.dev.
Favorite this tool to revisit it later, or Zap it to contribute to the public vote count.