Independent ML research
Mosaic Labs — Independent machine learning labIndex of small models · Vol. 00

Intelligence, piece by piece.

Small models. Focused intelligence.An independent lab exploring how machine learning can do more with less.

Explore the modelsRead the research
Directions
05
Released
00
Notes
Open
Fig. 01 — 128 pieces, no shared rule.
01 — Models

Small models. Big ideas.

An evolving collection of specialised models, built to explore what focused intelligence can achieve.

ExperimentResearchIn development
HoverTap to run
M/01 · Text → StructureIn development

Extract

Text in. Fields out.

Structured information extraction from unstructured text.

Explore Extract
M/02 · TextExperiment

Classify

Every sentence, sorted.

Lightweight text classification and categorisation.

Explore Classify
HoverTap to run
HoverTap to run
M/03 · RepresentationResearch

Embed

Meaning, as coordinates.

Compact semantic representations for similarity and retrieval.

Explore Embed
M/04 · VisionExperiment

Recognise

Learning to see, small.

Experimental computer vision and visual recognition.

Explore Recognise
HoverTap to run
M/06 · UnassignedOpen

The open tile

The next direction hasn't been chosen yet. Every mosaic keeps one space free.

Suggest an experiment
HoverTap to run
M/05 · CompressionResearch

Distill

Big model, small footprint.

Explorations in model compression and knowledge distillation: teaching small models what large ones already know.

Explore Distill
02 — Research

Not everything needs a billion parameters.

We explore the space between simple algorithms and massive foundation models — training smaller systems to do specific things exceptionally well.

Fig. 02 — The space betweenSchematic, not to scale
Hand-written rules
Classical ML
Small, specialisedMosaic
Large models
Foundation models
  1. 01

    Focused by design

    Models designed around specific problems.

  2. 02

    Built to understand

    Experimentation, training, and evaluation from first principles.

  3. 03

    Made to evolve

    An expanding collection of independent research experiments.

03 — Playground

See intelligence at work.

Pick a model, give it an input, inspect what comes back. A concept for the lab's live instrument.

Mosaic / PlaygroundPreview concept · example outputs, not live inferencePreview · not live
In — Sample textSample

Invoice INV-4471id from Northwind Ltdvendor is due on 12 Marchdate, total £1,240.00amount.

Out — Structured fields · JSONExample
{
  "invoice_id": "INV-4471",
  "vendor": "Northwind Ltd",
  "due_date": "12 March",
  "total": 1240.00,
  "currency": "GBP"
}
4 entities · illustrative example, not model output
Done · example outputLatency — · Params — · Not measured
04 — Journal

Notes from the lab.

Research experiments, engineering write-ups and technical discoveries, written up as we go.

Note 001Extract · EngineeringForthcoming

Building our first extraction model

From a blank repository to structured fields: data, labels, and a first honest evaluation.

002

Understanding embeddings from the inside

Embed · ResearchForthcoming003

When smaller models make sense

Essay · PhilosophyForthcomingAll notes
05 — Follow along

More intelligence.
Less excess.

An open-ended exploration of small models, thoughtful engineering, and the future of practical AI.

Explore the modelsFollow the research on GitHub
The open tile — hover to place the next piece