Your decisions
deserve your own
model.

LayaStudio fine-tunes open Laya decision models on your own labeled data, entirely on your Mac. No cloud, no per-call bill, nothing leaving the machine — and every run proves whether it actually got better.

Get it on GitHub ↗ See it play

git clone https://github.com/biplovgautam/LayaStudio && cd LayaStudio && uv run layastudioclone, install and open · Apple silicon

0tokens generated

36 msper decision

$0per call

8.8 msexported to Core ML

Base and fine-tuned models playing Snake side by side in LayaStudio — base vs fine-tuned, playing Snake unassisted on one Mac

What it does

four steps · one command

You bring labeled decisions, the studio does the rest: check the data, adapt the model, measure it honestly, hand you a checkpoint you can ship. Nothing leaves the machine at any step.

Read the docs ↗

Measured, not promised

apple m4 · 16 gb · balanced recipe · held-out rows

Task	Answers	Before	After	Training time

Test rows are never trained on, accuracy comes with a 95% interval and an exact McNemar test, and fine-tuning does not change inference speed: the adapters are merged into the weights.

The Snake run in LayaStudio: 15.8% to 98.8% move accuracy

Does it really learn?

snake, played unassisted

No hints and no safety layer: the model sees the board and four directions, its top answer is executed, and an illegal move ends the round.

	Moves	Apples	Legal
Base 322M	1.0	0.0	0%
Fine-tuned	169	19.8	99.3%
Planner (ceiling)	418	34.4	100%

Both public checkpoints are guessing at this task: 33.0% and 19.0% move accuracy, below the 25% a coin flip would give. Eighteen minutes of local fine-tuning takes it to 98.8%.

Then take it anywhere

every export verified against the runtime that trained it

Laya checkpoint

36 msFP16 safetensors with the original parameter names. Loads in laya-mlx on Apple silicon and in upstream PyTorch on Linux and NVIDIA — verified 40/40 identical answers.

Core ML · int8

8.8 ms308 MB for the Apple Neural Engine: four times faster than the runtime that trained it, a quarter of the size, and not one of 100 held-out decisions changed.

ONNX

83 msOpset 18, dynamic batch, tokens and options. Runs wherever onnxruntime runs — Linux and Windows CPU, CUDA, DirectML. Same 98.0% on the same held-out rows.

Every export is scored against the model it came from, on the same untouched rows — which is how you find out that int8 costs nothing here and int4 costs 28 points of accuracy.