colibri
Run frontier MoE models on hardware you already own, pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦
Runssource: GitHubhomepageCApache-2.0commit fd93c41aa6ae
C, Apache-2.0 licensed.
colibri runs. An Argusic agent installed it in 28 minutes and hit no errors on a clean machine with no GPU, and the whole session was recorded.
colibri v1.10.2 builds and installs from source in this container; 104 C test binaries (478 assertions) and 716 Python tests pass; the Kimi K3 engine produces token-exact greedy output matching the independent vendor oracle.
What the agent ended up with on a clean machine, in its own words. How this is measured
Measured by Argusic on a fresh machine every time. Every number links to its evidence. Argusic Score 73.3 of 100 (the mean of this project's run scores).
At a glance
Subject data from GitHub, linked at the top of this page, refreshed . Test data by Argusic (CC BY 4.0); every number links to a run page with the full log, the recording, and their sha256 hashes.
What was measured
From the run that decided this verdict. Every line is a recorded value, not an estimate.
- Install time
- 28 minutes
- Cold machine to finish
- 24 minutes
- Errors hit and fixed
- 0 hit, 0 fixed with no human help
- How the result was proved
- make test-c: ALL PASS (0 failures across 104 binaries, 478 ok assertions); make test-python: 716 passed, 59 skipped (GPU-only); c/tests/test_kimi_k3_tiny.py: token-exact greedy, teacher forcing, determinism, and oracle-bite checks pass on installed kimi_k3 binary; coli --version returns 'colibri 1.10.2'; coli info loads kimi_k3_tiny fixture reporting 'engine ready'
- Model tokens used
- 416,984
- Exact commit tested
- fd93c41aa6ae
- Container image
- sha256:cdd920bce783
What was not measured
The other half, so you can see exactly how far this evidence goes. A gap is printed as a gap and never filled in.
- Peak GPU memory
- this ran on a CPU machine, so there was no GPU to sample
- Behaviour over a long run
- each attempt uses the smallest documented settings and a fixed time budget, so this measures whether it runs, not how it holds up under load
The end of the recording
The last lines of the terminal session that decided this verdict, exactly as they were recorded.
**C test suite**: 104 test binaries, 478 "ok" assertions, **ALL PASS (0 failures)**. **Python test suite**: 716 tests, all pass (59 skipped for CUDA/Metal/Vulkan/GPU — expected in this… **End-to-end inference verified**: Token-exact greedy generation, teacher forcing, determinism, and… **No real credentials required**: the project needs no API keys, third-party services, or network t… tokens used 416,984 All tests verified. Here's the final report: **Install and build**: `make -C c` builds the GLM engine cleanly, plus `kimi_k3`, `olmoe`, `qwen36`… **C test suite**: 104 test binaries, 478 "ok" assertions, **ALL PASS (0 failures)**. **Python test suite**: 716 tests, all pass (59 skipped for CUDA/Metal/Vulkan/GPU — expected in this… **End-to-end inference verified**: Token-exact greedy generation, teacher forcing, determinism, and… **No real credentials required**: the project needs no API keys, third-party services, or network t…
Replay the whole session, every command from a clean machine to this point.
How it was tested
One agent, one container, no human help. It installs the project the way its own documentation says to, runs it, and fixes what breaks. Everything below is recorded as it happened: the terminal session, the log and the exact commit. The full method.
Strengths and limits
Measured facts, not opinions. How this is written.
What went well
- Reached a running state on a clean machine, with the session recorded.
- Installed in 28 minutes.
- Nothing broke on the way: zero errors between clone and running.
- Ran without a GPU, so it does not need one to start.
- Licensed Apache-2.0, as reported by its host.
- Measured 3 times, so the result is not a one-off.
What it is a reasonable choice for
- Trying it on a laptop or a small server: it reached a running state without a GPU.
- Following its own README: the documented path worked with nothing to fix.
Run history
| tested by | attempt | outcome | score | cost USD | when |
|---|---|---|---|---|---|
| Argusic Runner | 3/3 | Unverified | 60.00 | 0.16 | |
| Argusic Runner | 2/3 | Unverified | 60.00 | 0.14 | |
| Argusic Runner | 1/3 | Runs | 100.00 | 0.16 |
Embed the badge
Markdown for the project README. It links back here; terms on the terms page.
[](https://argusic.com/subject/colibri)Questions
- Does colibri run?
- Yes. colibri runs. Argusic installed and launched it on a clean machine in 28 minutes, hitting 0 errors on the way, and recorded the session.
- How did Argusic test colibri?
- On a fresh, disposable machine, with every command recorded and the repository pinned at commit fd93c41aa6ae. 3 attempts are recorded, and the full method is on the methodology page.
- What did it cost to test colibri?
- The run that produced this verdict cost $0.16: real compute and model cost, not a list price per million tokens. It is recorded on that run's page.
- How long does colibri take to install?
- 28 minutes in the run that decided this verdict, from a clean machine to a running state. That is one measured attempt, not an average.
- Does colibri need a GPU?
- Not to start. Argusic reached a running state on a machine with no GPU.
- What goes wrong when installing colibri?
- Nothing did, in the recorded run: zero errors between clone and running.
- Where is the evidence for colibri?
- All 3 recorded runs are on this page, each linking to its full log and terminal recording, stored with a sha256 fingerprint so it cannot be quietly altered.