⚗ AN AUTOMATE CAPTURE EXHIBIT · DISTILLATION, REFRAMED

Distill any task into a pocket-sized Spirit.

An open-source toolkit that turns a frontier teacher LLM into a 20–50M parameter model — a Spirit — that runs on CPU, edge, or browser. No API at inference. Pure model.

Production AI pipelines often chain six frontier LLM calls when only one of them needs to be frontier. Distillarium is a toolkit, a vocabulary, and a working reference Spirit for replacing the other five with tiny task-specific models you actually own.

See the Cellar → Read the docs →
v0.1.2 alpha Two bottled Spirits: Needle (function calling, 78% tool-name accuracy) and Routor (intent routing, 92% macro-F1). Research code, not production infrastructure — read each Spirit's Tasting Notes before you deploy.
TeacherGemini 2.5
Examples0
Proof0
Yield0 ml

What it is

Distillarium is three things in one repo.

① A TOOLKIT

Three commands — distill, taste, bottle — that take a YAML recipe, a teacher API key, and 30 minutes, and hand you back a deployable model.

② A VOCABULARY

Distillation has needed names for these concepts for years. We borrowed them from the people who've been doing this for centuries. Mash is your seed corpus. The Still is the training run. Cuts are your data splits. Proof is held-out accuracy. Tasting Notes are the eval report — strengths, weaknesses, and failure cases, auto-generated, no hiding.

③ A REFERENCE SPIRIT

Needle is a 20.7M-parameter attention-only transformer that does function calling. It was distilled from Gemini 2.5 Flash in 27 minutes for $0.30 on a single RTX 5090. It runs on CPU at ~45ms median latency. It is an alpha. The Tasting Notes say so honestly.

How it works

Write a recipe. Pour a mash. Run the still. Taste the cuts. Bottle what's good.

RECIPE

A YAML file naming your teacher, your task, and your student architecture.

MASH

Distillarium prompts the teacher to generate training data from a seed corpus.

THE STILL

The student trains on the teacher's outputs. Heads and tails are discarded; hearts are kept.

TASTING NOTES

Honest, auto-generated eval on held-out cuts. You see exactly what your Spirit can and can't do.

BOTTLING

Export to ONNX, GGUF, or browser-WASM. Run anywhere.

# Distill — uses GOOGLE_API_KEY from env
distillery distill recipes/needle.tool-calling.yaml

# Taste — auto-generates Tasting Notes against held-out cuts
distillery taste spirits/needle.pt

# Bottle — export to ONNX, GGUF, or WASM for deployment
distillery bottle spirits/needle.pt --format onnx

Every Spirit ships with Tasting Notes.

Most ML libraries report a single accuracy number and call it done. We auto-generate a full Tasting Notes report for every Spirit: tool-name accuracy, argument-key F1, exact-call accuracy, failure modes, retention vs. teacher (v0.2), and BFCL alignment for function-calling Spirits (v0.2).

The honesty is the brand. If a Spirit is 78% on tool names and 3% on exact calls, the Tasting Notes say so on the first line. You shouldn't have to read the source to find the weakness.

See Needle's Tasting Notes →

SAMPLE FROM NEEDLE'S NOTES
  • Strong tool-name selection on imperative queries (78%)
  • High arg-key precision (0.85) — never invents fake keys
  • Holds JSON structure consistently
  • Confuses adjacent social tools (Twitter vs Instagram)
  • Exact-call accuracy only 3% — values are the soft spot
  • WordPiece tokenizer adds whitespace noise (byte-BPE in v0.2)

🍾 The Cellar

A library of Spirits — what's bottled, what's aging, what's in the recipe book.

Bottled Ready to download, run, and inspect.
Aging Recipes drafted, partial distillations done, evaluation ongoing. Not yet released.
In the Recipe Book Ideas with recipes drafted, not yet distilled.

Explore the full Cellar →

What people are distilling

Four exhibits in the gallery — concrete scenarios where the technique earns its keep.

Intent routing on the edge

A customer chat front-end needs to classify "are you trying to buy, complain, or ask a question?" before routing to the right pipeline. A frontier LLM round-trip adds 800ms and $0.002 per call. An 8M Spirit (see Routor, currently aging) does it in 12ms for free, locally.

PII scrubbing before the cloud

Healthcare and legal apps need to strip PHI/PII from user input before anything reaches a hosted model. A 14M Spirit runs server-side at the trust boundary and never lets raw data leave — your compliance team gets an audit trail of a model you can actually inspect, not an API call to somebody else's box.

Tool routing for agentic systems

Your agent has 47 tools available. You don't need GPT-4 to pick which one to call — you need a 20M model that's seen the schemas and knows the patterns. Needle is the reference for this exact case: distilled from Gemini Flash on 15 categories, 78% tool-name accuracy, sub-50ms on CPU.

A repeatable thing you can audit

The Recipe is sealed — same teacher, same seed, same student arch equals the same Spirit. You can fork a public recipe, swap the tool catalog, and re-distill on your own data without ever calling a hosted API again at inference time.

🧭 How it compares

Distillarium sits in a specific gap in the existing distillation ecosystem. For breadth and SOTA function-calling, the alternatives below are excellent. For replacing a narrow LLM step with a CPU-runnable, audit-able artifact — that's our lane.

Arcee DistillKit
Sweet spot: general LLM distillation, 7B-target.
Distillarium targets the 5–50M class, deployment-first (.onnx / .gguf / .wasm as the output, not an afterthought).
ModelScope EasyDistill (incl. AgentKD)
Sweet spot: agent distillation, multi-modal.
We're CPU/edge-deployment focused, single CLI, Tasting Notes as the default eval rigor.
Berkeley TinyAgent
Sweet spot: 1.1B–7B function-calling SLMs.
We go smaller (20–50M) at the cost of generality — gain CPU inference and zero vendor lock-in.
LoRA fine-tuning
Sweet spot: capability extension on a generalist.
LoRA doesn't shrink the model. Distillarium produces a small student model, not a fine-tuned base.
Honest framing: for breadth and SOTA function-calling scores, TinyAgent is the right pick. For replacing one narrow LLM step in a production pipeline with a CPU-runnable artifact you can audit, fork, and ship — that's what Distillarium is for. BFCL numbers for Needle ship in v0.2 so the comparison is apples-to-apples.

🍾 Anatomy of a Spirit

Every bottled Spirit is a single self-contained artifact. Hover the parts to see what each one is in ML terms.

🟫
Cork — the Recipe. Sealed, signed config that produced this Spirit. You can re-distill from it byte-for-byte.
🧷
Neck — the inference shape. Tokenizer, vocab size, max sequence length.
🪟
Glass — the bottling format. PyTorch, ONNX, or GGUF wrapper around the weights.
🟧
Amber — the trained weights. The actual model parameters. Darker = higher proof.
📝
Label — the Tasting Notes. Name, vintage, batch, headline proof. Auto-generated, honest.

📜 The Vocabulary

Every term means something concrete. Each maps to an ML concept — and that's deliberate.

🌾Mashtraining corpus the teacher generates from
📜RecipeYAML config — versioned, forkable, reproducible
Cutstrain/val/test splits
🪜Heads · Hearts · Tailsfilter quality — kept / borderline / discarded
🔥The Stillthe training run itself
📈Proofheld-out accuracy. higher = more concentrated
📝Tasting Notesauto-generated eval report — strengths, weaknesses, samples
🛢Aging in Caskscontinued training / RLHF / refresh
🍾Bottlingexport — .pt / .onnx / .gguf / .wasm
🏛The Cellaryour (or the public) library of Spirits

Distillarium is an exhibit from Automate Capture — a creative-technical research studio. Built by Andrew Young. Distilled from candidates surfaced by Automate Capture's own Research Radar pipeline.

The CLI is free, runs locally, uses your own teacher API key, and produces models you own. Nothing reports back to us. The "Cellar" is a public showcase you can opt into. MIT licensed.

⚗ Talk to us about your pipeline

Tell us what your agent pipeline looks like today. We'll reply with a recipe sketch for distilling the narrow steps — usually within a day. No sales team, no SDR funnel.