[ VISUAL COMPOSER FOR TRANSFORMER ARCHITECTURE ]
Build a model.
Not a config file.
Otter is a visual composer for LLM transformer architectures. Drag attention, FFN, normalization blocks onto a canvas. See dimension contracts in real time. Export clean PyTorch you can train tonight.
[ MODEL ] llama-3-8b
[ B ] 1
[ S ] 2048
[ H ] 4096
[ HEADS ] 32 / 8
[ LAYERS ] 32
[ PARAMS ] 8.03B
[ 01 / COMPOSE ]
Drag, don't type
Eight components from real architectures. Llama, DeepSeek, PaLM. Connect them on a canvas and see dimensional compatibility before you run a single forward pass.
[ 02 / VALIDATE ]
Dimensions, live
Every node exposes its input and output contracts. The engine flags mismatches in 800 ms. Parameter count updates as you build.
[ 03 / EXPORT ]
Real PyTorch
Download a model.py, config.json, training script, and README. No lock-in, no runtime, no abstraction tax. Train it on whatever you have.
[ COMPONENTS — 8 SEEDED, 5 TEMPLATES ]
[ EMB ]
Token Embedding
[ ATTN ]
GQA Attention
[ FFN ]
SwiGLU FFN
[ NORM ]
RMSNorm
[ POS ]
RoPE
[ MOE ]
Mixture of Experts
[ OUT ]
LM Head
[ RES ]
Residual Add