100+ executable cards
Compose language, vision, multimodal and routing architectures from typed PyTorch atomics.
Loading Complexity
Loading Complexity
LABO AI turns PyTorch models into typed atomic graphs. Build by hand or ask an agent that uses the same visible, validated tools as you.
Free · Open source · macOS and Windows
Ask LABO
Build a compact GPT-like QA model and run it.
Token IDs
input
Embedding
200k → H
RMSNorm
H
Causal SDPA
8 heads / 2 KV
Residual MLP
SwiGLU
Residual add
H + H
LM head
H → vocab
One graph, four views
The visual graph, tensor contracts, generated PyTorch and local runtime share one typed model representation.
Compose language, vision, multimodal and routing architectures from typed PyTorch atomics.
Ask LABO can inspect, search, add, connect, arrange, run, save and export through bounded tools.
Stable ranks and lanes keep sequences, forks, joins and side-by-side comparisons readable.
Move between the visual graph and inspectable generated code without hiding the model behind a canvas.
Run, rerun, reset or step through an architecture and isolate failures between parallel models.
Save reusable presets locally, then export the complete graph as SVG or the model as Python.
Demo
Watch Ask LABO construct a compact GPT-like QA architecture, wire compatible ports, arrange it, generate PyTorch and execute it locally.
Open on YouTubeDownload v0.1.0
The PyTorch runtime stays on your machine. Add your own OpenAI API key only when you want to use Ask LABO.
Apple silicon · macOS 12 or later
Download DMGPortable ZIPWindows 10/11 · x64
Download installerPortable ZIPExplore the source, build a graph and tell us which atomic capability should come next.
View LABO AI on GitHub