Tiny LLM · Specification v1

Browser inference, with the model kept off the UI thread.

This milestone constructs the character-level decoder, performs causal forward propagation, produces logits, and generates text. It does not train or modify weights.

Inference only Worker-owned tensors 200,000 parameter maximum
01

Runtime visibility

Browser capabilities

Availability and active use are reported separately, so a detected API is never presented as an enabled accelerator.

WebGPU

Modern GPU compute interface.

Checking Waiting for the worker probe.

WebAssembly

Portable compiled browser runtime.

Checking Waiting for the worker probe.

CPU fallback

Specification-first float32 kernels.

Checking Waiting for the worker probe.
02

Execution boundary

UI remains responsive

Only commands, metadata, generated text, and requested logits cross the worker boundary.

03

Cross-engine contract

Python compatibility

The browser loads deterministic weights and expected outputs produced by the Python reference implementation.

Parity checks

Waiting
  • Tokenizer IDs
  • Parameter count
  • Tensor names and shapes
  • Manifest validation
  • Forward logits
  • Deterministic generation
Maximum logit error
—
Allowed tolerance
—

Loaded model

Loading
Architecture
—
Parameters
—
Context
—
Vocabulary
—
Blocks / heads
—
Tensors
—
Inspect canonical tensor table
TensorShapeParameters
04

Engine verification only

Untrained fixture check

This is not a trained language model. It verifies execution and Python compatibility using only the characters a, b, and space.

Compatible fixture prompt. Supported characters: a, b, and space.

Worker result 
Waiting for the model…