Open-Weight LLM · Private & Custom AI
sundial-base-128m
Purpose-built time-series foundation model for private forecasting automation—embed zero-shot predictions into ops workflows without external APIs.
Sundial is a 128M-parameter causal Transformer pre-trained on 1 trillion time points, designed for point and probabilistic time-series forecasting. For operations teams, it eliminates dependency on cloud forecasting services and enables in-house demand planning, inventory optimization, and anomaly detection. Deploy it on-premise to keep sensitive time-series data (sales, supply-chain, IoT sensor streams) fully contained.
Model facts
Private deployment
Run sundial-base-128m in your own environment
Self-host on modest CPU or GPU hardware (see estimates below). The model supports variable lookback/forecast windows and generates multiple probabilistic samples, enabling fine-grained forecasts within your private environment. No external API calls, no data egress. Trade-off: latency on large batch predictions (~500–950ms on M1 Pro, lower on A100); acceptable for batch or near-real-time ops workflows.
Operational AI use cases
Demand Forecasting & Inventory Automation
Feed historical sales/demand time-series directly into Sundial; generate 20+ probabilistic samples to compute confidence intervals and safety stock automatically. Integrate outputs into supply-chain workflows (reorder triggers, warehouse allocation) without sharing raw sales data externally.
Predictive Maintenance & Asset Monitoring
Ingest IoT sensor streams (temperature, vibration, pressure) from equipment; Sundial learns temporal patterns and flags anomalies or forecast degradation in performance. Alert ops teams in real-time, reduce unplanned downtime, all within your private infrastructure.
Financial & Revenue Forecasting for Planning
Automate weekly/monthly revenue, cash-flow, or expense forecasting from transaction/GL time-series. Generate quantile predictions (10th, 50th, 90th percentiles) for scenario planning and budget variance analysis—keep financial data private, accelerate board-ready forecasts.
Custom AI
As a base for custom AI
Use Sundial as a backbone for custom forecasting products or agents. Fine-tune on domain-specific time-series (energy load, network latency, churn signals) or wrap it in a workflow that combines forecasts with rules-based decisions (e.g., auto-scaling triggers, pricing adjustments). Generative sampling enables uncertainty quantification—crucial for downstream decision automation.
In the operating system
Where it fits
Sits in the **data/workflow layer** of a private AI OS: ingests clean time-series from ops databases (ERP, SCADA, billing systems) and outputs forecasts/confidence intervals fed into task agents, approval workflows, or direct automation triggers. No external LLM calls needed for time-series work, freeing budget for other AI initiatives.
Data control & security
Self-hosting eliminates data movement to external services. Time-series data (sales, supply-chain, IoT, financial) remains in your infrastructure throughout inference. This is an **architecture choice**, not a model property—you control where data lives, how it's encrypted, and audit trails. Compliance (HIPAA, SOX, GDPR) depends on your deployment, not the model.
Hardware footprint
**Estimate (FP32):** ~512 MB model weights + ~1–2 GB activation overhead per inference batch. On CPU (M1 Pro, 16GB): 249–950ms per forward pass depending on lookback/forecast/samples. On A100-40G GPU: sub-100ms. Suitable for on-prem servers with 4GB+ VRAM or multi-core CPUs; batch inference recommended for throughput.
Integration
Drop into Python/PyTorch via HuggingFace transformers library (v4.40.1+). Expects torch tensors of shape (batch, lookback_length); outputs tensor of shape (batch, forecast_length, num_samples). Wire via REST API (FastAPI, Flask) or batch jobs (Airflow, Prefect). Supports variable input lengths (672–2880 tokens); patch length = 16, adjust forecast horizon accordingly. KV cache & FlashAttention available for acceleration. Requires trust_remote_code=True (custom time-series code).
When it's not the right fit
- —Forecast horizon > 2880 steps (model context limit); will need sliding windows or coarser granularity.
- —Real-time ultra-low-latency (<50ms) requirements on CPU; GPU recommended but adds infra cost.
- —Time-series with extreme seasonal patterns or structural breaks not seen in pre-training; may need fine-tuning.
- —Irregular or very sparse time-series (e.g., rare events, long gaps); model expects regular temporal sampling.
Alternatives to consider
Chronos (Salesforce/amazon-chronos-1b)
1B-param time-series model, larger capacity but higher compute; same zero-shot paradigm, similar inference latency trade-offs. Pick Sundial for lower resource footprint.
TimeGPT (OpenAI/Nixtla, closed)
Cloud-only, API-based forecasting; no private deployment option. Sundial trades ease-of-use for full data sovereignty.
NeuralProphet (Facebook/Meta OSS)
PyTorch neural forecaster, smaller scope (no probabilistic sampling); requires more manual feature engineering. Sundial more out-of-box but heavier.
Related open models
FAQ
Can we fine-tune Sundial on our proprietary sales/demand data?
Yes—model supports fine-tuning via standard PyTorch. Load pre-trained weights, add your time-series, train on-prem with your own GPU/CPU. Exact code: see GitHub repo (thuml/Sundial). No API calls, data stays private throughout.
What's the commercial/business license situation?
Apache-2.0 licensed (permissive OSI). Free for commercial products, internal ops tools, and resale. No royalties, no restrictions. Just include Apache notice. Confirm with legal for your use case, but no red flags.
How do we integrate forecasts into our ERP or BI tool?
Expose Sundial as a microservice (FastAPI/Flask) running in your environment. ERP/BI calls REST endpoint with time-series batch; returns (forecast_length, num_samples) tensor. Parse samples for mean, quantiles, confidence intervals. Latency ~1–2 sec per batch on CPU; design accordingly.
Does Sundial handle multiple time-series (e.g., forecast 1000 SKUs in parallel)?
Yes—batching is built-in. Shape (1000, 2880) → (1000, 96, 20) (1000 SKUs, 96 forecast steps, 20 samples). Batch size limited by VRAM/CPU memory. A100 can handle large batches; CPU works fine for smaller batches. Use data loader / batch jobs for scale.
Ready to Build Private Forecasting Automation?
Sundial is purpose-built for on-prem time-series work. LLM.co helps you integrate it into your ops stack—data stays yours, forecasts stay private. Let's architect a custom AI system for your planning, monitoring, or automation needs.