Colibri explained: streaming MoE experts across storage, RAM and VRAM
Colibri quickstart: check capacity and diagnostics before chatting
Use the documented source path without accidentally starting a large model download.
What you will learn
- Check the chosen model first
- Build or select a matching release
- Diagnose before generating
Before you start
- Basic Python and command-line knowledge
- A documented hardware and storage inventory
Create a reproducibility ledger before attempting another inference optimization.
Key takeaways
- Capacity estimates are model-specific.
- Old archive examples are not latest releases.
- Diagnostics come before model evaluation.
Check the chosen model first
The quickstart’s approximate 16 GB RAM and 380 GB disk figures describe its GLM-5.2 int4 example, not a universal requirement. Other containers differ; budget temporary conversion space and preserve free space for the operating system.
Choose the model and container deliberately and review its license before downloading. The article does not run convert or obtain weights automatically. A ready-made container and conversion from source weights are separate preparation routes.
Build or select a matching release
For source builds, the documented flow enters colibri/c and runs setup.sh after installing the appropriate compiler, make and OpenMP support. Pinning the source below keeps the reading reproducible; inspect the script before execution.
Prebuilt archives are another documented route, but availability depends on platform and release. Do not copy the old v1.1.0 archive example as a current release recommendation. Record the actual archive and architecture selected.
Diagnose before generating
After preparing a supported model directory, use coli doctor to inspect readiness and coli plan to inspect placement. Start chat only after checking model location, permissions and available memory.
The quickstart reports a small engine self-test with toolchain-dependent floating-point near ties. That is upstream guidance, not a result obtained here. Investigate unexpected failures rather than treating a partial download as a model-quality problem.
Decision guide
| Criterion | Option A | Option B |
|---|---|---|
| Best when | You need predictable behavior and easy auditing | You need adaptive optimization and have reliable telemetry |
| Main risk | May leave performance on the table | Can become difficult to explain or debug |
Implementation steps
- 1
Budget the selected container and temporary space.
- 2
Review and build a fixed source revision or matching archive.
- 3
Run diagnostics and placement review before chat.
Copy-ready example
git clone https://github.com/JustVugg/colibri.git
cd colibri
git checkout 9d5d05de7f4ccf39840224ed295f292ab8aeb598
cd c
# Review setup.sh and prerequisites before running:
./setup.sh
# After separately preparing licensed model files:
COLI_MODEL=/path/to/model ./coli doctor
COLI_MODEL=/path/to/model ./coli planFrequently asked questions
Does this command download a model?
The source checkout/build example does not include the separate model download or conversion command.
Was setup.sh executed here?
No. The procedure is documented from fixed upstream sources.
Sources
- Colibri / docs/quickstart.mdSource checked 2026-09-23