Sundial (THUML) — GGUF

GGUF conversion of THUML's Sundial — a 128M parameter flow-matching generative model for time series. Converted and run with zsfm, a Rust workspace that ports zero-shot forecasting and tabular foundation models to GGUF + candle. No PyTorch, no Python runtime required to run inference.

F16 is generally the best size/accuracy trade-off; Q8_0 is smallest. This repo's default recommendation matches the upstream conversion default: F16.

Context must be at least 32 timesteps (Sundial pads short contexts to a full patch, but 32 gives a meaningful example) — a shorter context fails with context too short. The examples below use a 32-value context.

Use it

Python (pip install zsfm)

pip install zsfm huggingface_hub

import zsfm
from huggingface_hub import hf_hub_download

gguf_path = hf_hub_download("amaye15/sundial-gguf", "sundial-f16.gguf")

model = zsfm.SundialModel(gguf_path)

context = [0.85, 0.93, 1.01, 1.09, 1.17, 1.25, 1.33, 1.06, 1.14, 1.22, 1.3, 1.38, 1.46, 1.54, 1.27, 1.35, 1.43, 1.51, 1.59, 1.67, 1.75, 1.48, 1.56, 1.64, 1.72, 1.8, 1.88, 1.96, 1.69, 1.77, 1.85, 1.93]
point = model.forecast(context, horizon=64)
# -> List[float], the (deterministic) decoded point forecast

Sundial is generative (flow-matching): this Rust port's ODE solver starts from an all-zero state rather than sampled noise, so forecast() is deterministic — a single decoded path, not a distributional sample or an ensemble mean.

Rust / CLI (cargo install zsfm)

cargo install zsfm --locked

# downloads the original weights and converts to GGUF locally
# (produces the same bytes as sundial-f16.gguf in this repo):
zsfm sundial convert --dtype f16 -o gguf/sundial-f16.gguf
echo '{"context": [0.85, 0.93, 1.01, 1.09, 1.17, 1.25, 1.33, 1.06, 1.14, 1.22, 1.3, 1.38, 1.46, 1.54, 1.27, 1.35, 1.43, 1.51, 1.59, 1.67, 1.75, 1.48, 1.56, 1.64, 1.72, 1.8, 1.88, 1.96, 1.69, 1.77, 1.85, 1.93], "horizon": 64}' \
  | zsfm sundial infer --gguf gguf/sundial-f16.gguf

-m/--model takes the full HuggingFace repo id (default thuml/sundial-base-128m) — there's only one published checkpoint for this architecture, so you normally don't need to change it. -o/--output defaults to gguf/sundial-f16.gguf regardless of --dtype, so always pass -o explicitly (as above) — otherwise repeated runs overwrite the same file under a name that may not even match the dtype you chose:

zsfm sundial convert --dtype f32 -o gguf/sundial-f32.gguf
zsfm sundial convert --dtype q8  -o gguf/sundial-q8.gguf

To skip conversion and run a file already published here:

huggingface-cli download amaye15/sundial-gguf sundial-f16.gguf --local-dir .
echo '{"context": [0.85, 0.93, 1.01, 1.09, 1.17, 1.25, 1.33, 1.06, 1.14, 1.22, 1.3, 1.38, 1.46, 1.54, 1.27, 1.35, 1.43, 1.51, 1.59, 1.67, 1.75, 1.48, 1.56, 1.64, 1.72, 1.8, 1.88, 1.96, 1.69, 1.77, 1.85, 1.93], "horizon": 64}' \
  | zsfm sundial infer --gguf sundial-f16.gguf

Source, the other 9 time-series forecasters + 5 tabular models, and full docs: amaye15/zsfm-rs.

Response format

{
  "id": "forecast-000001932b7a1234",
  "object": "forecast",
  "created": 1736290000,
  "model": "sundial",
  "choices": [{
    "index": 0,
    "forecast": {
      "point": [2.1, 2.3, 2.5],
      "quantiles": {}
    },
    "finish_reason": "stop"
  }],
  "usage": {"context_length": 32, "forecast_length": 64}
}

This architecture has no quantile head, so quantiles is always empty and point is the only forecast.

Pass a batch of series ("context": [[...], [...]]) for one choice per series.

Architecture

Sundial is a flow-matching generative model for time series:

• Encoder: Patch-based causal transformer (128M params) that encodes the context window into a conditioning vector
• Decoder: A small flow network trained to map Gaussian noise to the forecast distribution, conditioned on the encoder output
• Sampler: This port runs Euler integration from an all-zero initial state (not sampled noise) for a deterministic, reproducible decode
• RoPE: Standard rotary positional embeddings applied to all attention heads

License

Conversion code: MIT (amaye15/zsfm-rs). Weights: Apache-2.0, per THUML's original release — unrestricted, including commercial use.