ReleasedOpen Source • Python • 2026v0.1

Distillarium

A toolkit for distilling small, task-specific models from teacher-generated data.

The released toolkit includes Needle, a 20.7M-parameter reference model for a narrow tool-calling task, plus training, evaluation, and export workflows.

Open site ↗ Source ↗ PyPI ↗

Reported facts

Reference Spirit: Needle (20.7M)
Reported training run: 27 min · 1× RTX 5090
Reported teacher API spend: $0.30
Reported tool-name accuracy: 78%
Reported argument-key F1: 0.73
On-disk size: 249 MB
Reported CPU latency: ~45 ms median

Quickstart

# 1. Install
pip install distillarium[gemini]

# 2. Distill — uses the reference Needle recipe
distillery distill recipes/needle.tool-calling-v1.yaml

# 3. Bottle the Spirit for deployment
distillery bottle spirits/needle.pt --format onnx

The published Needle reference run reports about 27 minutes on one RTX 5090 and $0.30 in teacher API spend. Results depend on the recipe, teacher, hardware, and task.

A narrow reference result

Needle demonstrates one task-specific distillation recipe. It is evidence for that configuration, not a claim that every task can be compressed to the same size or quality. The repository and package provide the implementation for inspection and further testing.

Project vocabulary

Spirit: The trained, bottled model — the output of a run.
Mash: Seed corpus the teacher generates training data from.
Recipe: YAML config: teacher, mash, student arch, cuts, still, tasting, bottling.
The Still: The training run itself.
Cuts: Train / eval / test data splits.
Heads / Hearts / Tails: Discarded noise / kept core / borderline cases.
Proof: Held-out accuracy. The higher the proof, the more concentrated.
Tasting Notes: Auto-generated eval report — strengths, weaknesses, failure cases.
Aging in Casks: Continued training, fine-tuning, RLHF refresh.
Bottling: Export to ONNX, GGUF, or browser-WASM.
The Cellar: Library of Spirits (public or private).

Provenance

How this work began

The project grew from experiments on whether a small, task-specific model could reproduce a limited tool-calling behavior. Distillarium packages that workflow and publishes the reference implementation.

The Research Radar Protocol is a proposed specification. This record does not claim an on-chain attestation.