Crucible
Compresses large MoE language models into compact specialists for local inference.
Tested on LinuxKnown vulnerabilities
Use it
Before you install: known vulnerabilities
pickfit found 1 known high or critical vulnerability in the packages a fresh install brings (OSV.dev, 3 Nov 2026). Weigh this before you install it.
pypi:certifi 2026.2.25certifi bundles emSign root CAs with compromised intermediate CA keys (GHSA-6q48-6c7w-9gw2, critical, fixed in 2026.10.18)
Linux x86_64
You need first
- Runtimes:
python 3.12+
clone repo
git clone --depth 1 https://github.com/peonist-ai/crucible "$HOME"/crucibleinstall via uv sync
cd "$HOME"/crucible && uv sync
Where the README and reality differ
- README: uv pip install crucible-bench installs the benchmark harness
What worked: clone https://github.com/peonist-ai/crucible then uv sync from repo root; PyPI crucible-bench 1.6.0 is a different thesis-judgment tool
Verified in pickfit: the commands exactly as its sandbox ran them
For the record only: they name the sandbox's home, /home/sbx.
git clone --depth 1 https://github.com/peonist-ai/crucible /home/sbx/cruciblecd /home/sbx/crucible && uv synccd /home/sbx/crucible && uv run crucible --helpCheck that it works
cd "$HOME"/crucible && uv run crucible --helpThe output should contain “MoE expert compression toolkit”.
What it does
- Prune experts from MoE modelsNot tested
- Merge experts, then prune themNot tested
- Quantize pruned models to W4A16Not tested
- Benchmark served modelsNot tested
How far pickfit got
Linux x86_64
- Installs
- Starts
- 0 of 4 features work
- Works offline
Good for
- score MoE experts on calibration data
- prune 40% of experts with REAP
- quantize pruned model to W4A16
- benchmark served model via OpenAI-compatible endpoint
- compare baseline and compressed benchmark runs
How pickfit tested it
Each run starts from a clean machine and follows the steps above. How pickfit verifies
Linux x86_64
Installs and starts
- Replayed from scratch
- Works offline
- Checked on 30 Oct 2026
- Version 0.1.0