Shrink a real model with your own hands, and watch what it costs.
Pigeon is a small image model with one job: is there a bird, and where? Quantize it, prune it, palettize it, right here in your browser, on 500 real photos. Keep it accurate and it slims from a rock pigeon to a common swift. Push further with these tools, which never retrain it, and it starves.
The Coop
Where the shrinking happens. Stack quantization, pruning and palettes into a recipe. After every change Pigeon scores itself again on all 500 photos and shows what it cost: accuracy, pointing, and the real size in bytes. The Coop's second cage holds YOLOv8n, a real detector that draws a box around every bird it finds. Its tools are its own: one precision for all its weights, and a channel cut for each stage. It is scored on the same 500 photos, and its health is the share of its bird-finding AP it keeps.
Rock Pigeon
Columba livia
uncompressed350 g
Mourning Dove
Zenaida macroura
≥ 1.8× smaller120 g
Common Swift
Apus apus
≥ 3.5× smaller40 g
Barn Swallow
Hirundo rustica
≥ 7× smaller18 g
No healthy Pigeon this small yet
Ruby-throated Hummingbird
Archilochus colubris
≥ 14× smaller3 g
No healthy Pigeon this small yet
Skeleton
pushed too far:it's guessing
The key to sizes: how much smaller a model gets decides which bird it becomes. Pigeon stays healthy while it keeps ≥ 90% accuracy and ≥ 75% pointing (its heatmap lands on the real bird). Below that it starves, and it is bones at ≤ 60% accuracy or ≤ 35% pointing.
The smallest healthy Pigeon on record is a Common Swift: the Mixed precision preset, 737 KB, 5.00 times smaller, with 94.4% accuracy and 98.0% pointing (recorded in the lab's notes, not measured on this page). No recipe on record keeps it healthy at 7 times smaller or more, where the Barn Swallow starts. That probably takes retraining, which the Coop's tools never do; the Field Guide's Quantization-aware training and Knowledge distillation (Hinton) show how it works on the lab's small models.
Every bird you hatch, in either cage or in a Field Guide run, is kept on your life list.
Open the Coop with a preset loaded
- FP16Half-precision floats. Almost always free.
- INT8, one scaleOne scale for a whole tensor. Watch what happens.
- INT8, per channelA scale for every output channel.
- Mixed precisionSqueeze-excite at 4-bit, everything else at 8-bit.
- MXFP44-bit floats. Each block of up to 32 weights in an output channel shares a power-of-two scale.
- 2:4 sparsityKeep the biggest 2 of every 4 weights.
- 16-colour palettek-means: every weight becomes one of 16 shared values.
Lessons, methods, experiments and your birds
- Flight School10 coding lessonsWrite the core of a method yourself in Python, then check it against tests, all in your browser. Plus a 12-chapter course on the building blocks that reads through the Field Guide; 6 of its chapters still have gaps, and each says what is missing.
- Field Guide43 methodsOne write-up per method, grouped by what it does: distillation, pruning, quantization, factorization, training small models, specialization and depth reduction. 31 of them run in your browser.
- Results28 experimentsSimulations in 6 groups, each asking one question about compression, with its figures, its numbers, and a report on what they show.
- Life listkept in your browserEvery bird you hatch in the Coop's two cages and in Field Guide runs, the way a birder keeps a life list: each species once, at its best sighting, with the recipe that made it.