leda/language model

Playground / story continuation

A language model trained from scratch. Give it a beginning.

Runs on your device
Edit the opening, or try the one above.Loading tokenizer…
Parameters256-token context

Draw from the model’s distribution.

Lower is steadier. Higher is more varied.

Includes the end token.

Repeatable within this runtime.

Context0 / 256
0 prompt0 generated256 remaining
First use downloads 63 MB · Cached on this device
02 Continuation0 TOKENS

The continuation will appear here.

Click any generated token to inspect it or choose a different path.

Leda · 15.7M parameters · Built by Brandon Ling

Trained from scratch in C++ with Spar ↗ on GPT-4-generated TinyStories.

A small story-continuation model, not a chat assistant. It can repeat itself, mix up characters, or lose the plot.

About the model

Six layers, width 384, six attention heads, SwiGLU, RMSNorm, QK normalization, RoPE, and tied embeddings. Trained on 2,707,825 prepared stories over 67,960 updates. This is the final September 8, 2026 checkpoint, with a strict 256-token context.

Full Float32 inference uses WebGPU when available, with a WebAssembly CPU fallback. Speed depends on your device. Seeds reproduce a run within a runtime; small numerical differences between devices can change choices.

Model details & inference instructions ↗