04 / INFERENCE SPIKE

Native WebGPU text model

Pure JavaScript and WGSL. No inference libraries.

MODELGlassbox GPT · Tiny Shakespeare · 10.8M parameters · 65 character vocabulary · 256 character context · MITChoose an inference target

WEBGPU CHECK PENDING

MODEL TARGET

This small research model continues Shakespeare-like text. It does not follow instructions and is not suitable for NPC dialogue. Its job here is to test a hand-written WebGPU transformer runner.

Weights download only when you request them; prompts stay in this browser.

Model storage status will appear when you load a model.

STARTING TEXT

NEW TOKENS TEMPERATURE TOP K REPETITION PENALTY

GENERATED TEXT

Load the model, then generate a short continuation.

Inference: —Generated: 0 tokens

How this spike is built

The page reads a local safetensors file and runner config, uploads weights to WebGPU storage buffers, and runs embedding, normalization, causal attention, feed-forward, and sampling using project-owned WGSL and JavaScript. No external JavaScript or inference library is loaded.

Model: Glassbox GPT — Shakespeare, MIT License. It was trained on the Tiny Shakespeare corpus. Its model card documents that its output is stylistic mimicry and often nonsensical.