idle-intelligence/t0-beta-q4_0-webgpu

t0-beta Q4_0, WebGPU

Q4_0-quantized weights for theforecastingcompany/t0-beta, packaged for client-side browser forecasting via WASM + WebGPU.

Runs entirely in the browser, no server required. Probabilistic multi-horizon time-series forecasting, 256M parameters, smallest/fastest quant in this release.

**Try the demo →**

What you gain, what you lose

You trade more accuracy than on t0-alpha for a much smaller file: 149.5 MB, 0.14x the F32 weights. Point drift worst-case against our F32 reference is 14.6%, looser than t0-alpha's own Q40 (8.4%) and looser than either quant level of this checkpoint should be for production use. There is no full-97-config GIFT-Eval run for t0-beta at any quant level, only the 8-config subset below. Use Q80 for this checkpoint unless file size is the binding constraint.

Files

Usage

These weights are consumed by t0-web, a Rust/WASM + WebGPU forecasting engine built with Burn.

await t0wasm.initBackend();
const modelBuf = await fetch('t0-beta-q4_0.gguf').then(r => r.arrayBuffer());
const model = t0wasm.T0Wasm.load(new Uint8Array(modelBuf));

const context = series.slice(-512);
const quantiles = await model.forecast(context, 32);

Weights are fetched from this repo and cached by the browser.

Requirements

• —Chrome 113+ or Edge 113+ (WebGPU required)
• —HTTPS (required for WebGPU)
• —~150 MB download on first load (cached afterward)

Pipeline

Series → patches of 32 (96-vector each)
  → 24 transformer blocks [WASM, WebGPU] → time and group attention, embed 1024
    → 32-step quantile decoder → 21 quantile levels
      → autoregressive rollout for longer horizons

Benchmarks

Drift vs our own F32 reference, and vs the official published t0-beta INT8 card

Q40's point drift exceeds the official published INT8 card's on this checkpoint. Wider `embeddim` (1024 vs t0-alpha's 512) gives Q80's per-channel blocks more values to average error over, but Q40's fixed 32-value block granularity does not scale with that width the same way.

GIFT-Eval, official-protocol 8-config subset (dequantized back to f32 into the reference architecture)

Within 0.4% relative of the f32 reference on this small subset, a looser signal than the drift table above. This 8-config subset is not comparable to the published 97-config headline numbers (CRPS 0.4738 / MASE 0.6865); no full-97-config run exists for this checkpoint.

Latency (native Metal only, no browser measurement)

Measured with t0-fast on raw wgpu/Apple Metal, context 512, horizon 32. No headless-Chromium browser run exists for t0-beta; do not read this as a browser latency figure.

Model Details

• —Base model: theforecastingcompany/t0-beta by The Forecasting Company
• —Architecture: Patch transformer, time and group attention
• —Parameters: ~256M
• —Quantization: Q4_0 for attention.wQKV.weight, attention.wO.weight, mlp.0.weight, mlp.2.weight per layer; norms, embeddings, biases, and the quantile head kept at f16
• —Quantile levels: 21
• —License: Apache-2.0 (same as original)

Quantization

Weights-only quantization using standard GGUF Q4_0 blocks (32 values per block, fp16 scale), in ggml-compatible layout, dequantized on-GPU inside the WGSL matmul with F32 compute. Exported from the F32 safetensors by t0-web's own packer. The F32 path itself matches the PyTorch reference to 3.2e-6 max-abs.

Citation

@misc{tfc-t0,
  title  = {t0: A time-series forecasting foundation model},
  author = {The Forecasting Company},
  year   = {2026},
  url    = {https://huggingface.co/theforecastingcompany/t0-beta},
}

Disclaimer

This is an independent port by ilnmtlbnm@idle-intelligence, not affiliated with or endorsed by The Forecasting Company. Forecast values may differ slightly from the original PyTorch implementation due to quantization.