index.mdx

title	colibrì documentation
description	Run frontier mixture-of-experts models across storage, RAM, and VRAM.

colibrì is a dependency-light inference engine for running large mixture-of-experts models on consumer and heterogeneous hardware. It treats storage, RAM, and VRAM as one inference hierarchy while keeping model semantics explicit and measurable.

Build the engine, prepare a model, and start an interactive chat. Serve OpenAI- and Anthropic-compatible endpoints from one persistent model process. Browse request schemas, responses, authentication, and live examples by endpoint. Measure and tune placement across VRAM, RAM, and storage without silently changing precision.

What you can run

The repository currently includes engines for GLM, DeepSeek, Kimi, Qwen, Inkling, and OLMoE families. Each engine uses the same coli chat, coli serve, and coli web front end.

colibrì is an inference engine and research platform, not a hosted service. Throughput depends heavily on model format, memory placement, storage, and prompt size.

Common entry points

Command	Use it for
coli chat	Interactive terminal inference
coli serve	OpenAI- and Anthropic-compatible HTTP endpoints
coli web	HTTP endpoints plus the local web dashboard
coli doctor	Environment and dependency diagnostics
coli tune	Measured placement and performance tuning

Set `COLI_API_KEY` before exposing the HTTP server beyond a trusted local machine. Keep raw browser and telemetry endpoints behind the same network boundary.