Early Access — Otter v0.1

Build transformer architectures with visual precision.

Compose ideas from modern research into a custom model. The canvas validates compatibility in real time. Export clean PyTorch.

Start buildingRead docs
my-custom-model.otter
Embedding
Vocab: 128k
GQA Attention
Llama-3 (8 Heads)
SwiGLU FFN
Dim: 14336
LM Head
Softmax Out
TRANSFORMER BLOCK [ x32 LAYERS ]
Valid - 8.1B params

Visual Composition

Drag and drop layers visually. Validate dimensional compatibility in real time before running a single forward pass.

Battle-Tested Components

Mix and match GQA, MoE routing, and SwiGLU directly from architectures like Llama-3 and DeepSeek-V3.

Clean PyTorch Export

Export pure, documented PyTorch code. No hidden abstractions, no lock-in. The architecture is yours to train.