DEV.co
Open-Source LLM · thuml

sundial-base-128m

Sundial is a 128M-parameter generative time-series foundation model pre-trained on 1 trillion time points. It performs zero-shot point and probabilistic forecasting by generating multiple plausible future sequences. The model uses a decoder-only Transformer with a TimeFlow loss function to produce non-deterministic predictions, enabling statistical analysis (quantiles, confidence intervals) from sample sets rather than single-point estimates.

Source: HuggingFace — huggingface.co/thuml/sundial-base-128m
128M
Parameters
apache-2.0
License (OSI-approved)
Unknown
Context (tokens)
148.4k
Downloads (30d)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

FieldValue
Developerthuml
Parameters128M
Context windowUnknown
Licenseapache-2.0 — OSI-approved
Modality / tasktime-series-forecasting
Gated on HuggingFaceNo
Downloads148.4k
Likes78
Last updated2026-03-09
Sourcethuml/sundial-base-128m

What sundial-base-128m is

Sundial is a causal Transformer (decoder-only, 12 layers, FP32) designed for univariate time-series forecasting. It patches input sequences (patch length 16), embeds them, and applies auto-regressive decoding with TimeFlow loss—a flow-matching-based objective that models per-token probability distributions. Supports lookback up to 2880 steps, multi-patch prediction up to 720 steps, and generates multiple samples for probabilistic forecasting. Includes optimizations (KV Cache, FlashAttention, FlashAttention-compatible). Pre-trained on 1032B time points from public datasets (UTSD, Salesforce/lotsa_data, autogluon/chronos_datasets).

Quickstart

Run sundial-base-128m locally

Load the open weights with 🤗 Transformers and generate — the same model, self-hosted.

quickstart.pypython
from transformers import pipelinepipe = pipeline("text-generation", model="thuml/sundial-base-128m")out = pipe("Explain retrieval-augmented generation in one sentence.",           max_new_tokens=128)print(out[0]["generated_text"])

Swap in vLLM or Ollama for production-grade serving. DEV.co can stand up the inference stack.

Deployment

How you'd run it

A typical self-hosted path — open weights, an inference server, your application.

DEV.co builds each layer — from GPU infrastructure to the application.

Best use cases

Probabilistic Forecasting with Uncertainty Quantification

Generate multiple plausible future trajectories to derive confidence intervals, quantiles, and full distributional forecasts—ideal for risk-aware planning in finance, supply chain, or energy.

Zero-Shot Forecasting Across Domains

Leverage pre-training on 1 trillion time points to forecast new time-series domains without fine-tuning. Card shows top-1 performance on GIFT-Eval and Time-Series-Library benchmarks.

Real-Time IoT and Streaming Time-Series

Inference on Apple M1 Pro CPU achieves ~500ms for lookback 2880 + forecast 16; integrates with Apache IoTDB for production time-series databases.

Running & fine-tuning it

128M parameters, FP32 precision. VRAM estimate: ~512 MB for model weights + ~1–2 GB working memory (batch inference, KV cache). CPU inference feasible (M1 Pro ~500ms); A100 GPU provides 4–8× speedup. Transformers 4.40.1 + Python 3.10 required for stability.

Card does not document fine-tuning procedures, LoRA, or QLoRA feasibility. Model includes custom_code tag; trust_remote_code=True required for loading. Fine-tuning approach unknown—requires review of GitHub repo (thuml/Sundial) or paper (arxiv:2502.00816) for guidance. Decoder-only architecture theoretically supports parameter-efficient tuning, but no official instructions provided.

When to avoid it — and what to weigh

  • Multivariate or High-Dimensional Time-Series — Card does not explicitly state multivariate capability. Use case appears limited to univariate sequences; review paper (arxiv:2502.00816) or notebook for multi-channel support.
  • Very Long Context Windows — Maximum context length is 2880 steps. If your forecasting task requires lookback windows larger than this, architecture is not suitable without re-training or aggregation strategies.
  • Sub-millisecond Latency Requirements — Inference on M1 Pro is 500ms+ for typical workloads. GPU (A100) is faster but still not ultra-low-latency; unsuitable for real-time control or high-frequency trading.
  • Deterministic Single-Point Predictions Only — Model is designed for generative sampling. If you need only a single deterministic forecast, added complexity of sampling and aggregation may be unnecessary overhead.

License & commercial use

Apache-2.0 license. Permissive OSI-approved license allowing modification, distribution, and commercial use, provided original copyright and license notice are retained.

Apache-2.0 is a permissive OSI license that explicitly permits commercial use. No gating or restrictions stated. However, review your use case against Apache-2.0 terms (warranty disclaimers, liability limits). No commercial support SLA or indemnification mentioned; consult Legal if production deployment requires SLA.

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

SignalAssessment
MaintenanceActive
DocumentationAdequate
License clarityClear
Deployment complexityModerate
DEV.co fitStrong
Assessment confidenceHigh
Security considerations

Model uses trust_remote_code=True for custom_code loading; review source before deployment in restricted environments. Pre-trained on public internet data (UTSD, Salesforce, autogluon datasets)—inherited data quality/bias risks unknown. No explicit threat model, adversarial robustness, or data provenance audit stated. Consider input validation and output monitoring for production time-series systems.

Alternatives to consider

Chronos (Amazon/Salesforce)

Similar 128M architecture for zero-shot time-series forecasting. Cited in card (arxiv:2403.07815). Compare training scale, inference speed, and multivariate support for your use case.

Time Series Transformer (Hugging Face / Google)

Established baseline for univariate forecasting. Less recent but mature documentation and broader framework support (JAX, TensorFlow, PyTorch). No probabilistic generation.

NeuralProphet or StatsForecast

Lighter-weight, interpretable alternatives for time-series forecasting. Lower inference cost; no generative sampling but faster deployment in resource-constrained environments.

Software development agency

Ship sundial-base-128m with senior software developers

Sundial is open-source (Apache-2.0) and ready for production. Start with the quickstart code, benchmark against your data, and evaluate on GIFT-Eval or Time-Series-Library. Contact the authors or consult the GitHub repo for fine-tuning guidance.

Talk to DEV.co

Related open-source tools

Surfaced by semantic similarity across the DEV.co open-source index.

sundial-base-128m FAQ

Can I use Sundial for commercial applications?
Yes. Apache-2.0 explicitly permits commercial use, provided you retain copyright and license notices. No gating applied. However, no SLA or commercial support is mentioned; contact authors or check GitHub for support options.
What hardware do I need to run Sundial?
Minimum: 2 GB RAM for inference on CPU (M1 Pro ~500ms per inference). GPU (A100) speeds up 4–8×. For production, A100 or equivalent recommended. Requires Python 3.10, transformers 4.40.1.
Does Sundial support multivariate time-series?
Card and quickstart show univariate examples only. Multivariate support is not clearly stated. Review paper (arxiv:2502.00816) or GitHub repo for details; email authors if critical to your use case.
How do I fine-tune Sundial on my own data?
Card does not provide fine-tuning instructions. Check GitHub (thuml/Sundial) or paper for guidance. LoRA/QLoRA feasibility unknown. Contact authors at [email protected] or [email protected] for training scripts.

Software developers & web developers for hire

Need help beyond evaluating sundial-base-128m? DEV.co is a software development agency offering software development services and web development for teams of every size. Our software developers and web developers build custom software, web applications, APIs, and open-source llms integrations — and maintain them long-term.

Ready to Deploy Time-Series Forecasting?

Sundial is open-source (Apache-2.0) and ready for production. Start with the quickstart code, benchmark against your data, and evaluate on GIFT-Eval or Time-Series-Library. Contact the authors or consult the GitHub repo for fine-tuning guidance.