NVIDIA Nemotron Labs Dory1.0-4B
A downloadable English reasoning model with adjustable inference loops, long context, and local tool-calling APIs.
Timestral
github.com
Downloadable forecasting models turn numerical time-series history and covariates into multivariate predictions with uncertainty estimates.
Visit website
Pricing
Free
Available on
Python, PyTorch CPU, PyTorch GPU
Category
Developer Tools
What is Timestral?
Timestral is Mistral AI's family of downloadable time-series forecasting models. It produces probabilistic forecasts from numerical histories and can incorporate related variables known in the past or future. It serves data scientists and developers building forecasting systems for capacity planning, energy, supply chains, and equipment monitoring.
The family supports zero-shot forecasting of individual series or multiple related variables. Four checkpoints offer different model sizes, while a standalone Python library provides local inference on CPU or GPU.
Key features
Joint multivariate forecasting
Timestral forecasts related target variables together and also supports univariate inputs, allowing the same family to address different forecasting structures.
Probabilistic outputs
Predictions include 21 quantiles per time step. Applications can retrieve the median and selected quantiles to represent a range of possible outcomes.
Past and future covariates
Inputs can include past-only, known-future, and categorical covariates. Categorical values use numeric codes, and covariates condition the target forecasts.
Four downloadable sizes
Choose checkpoints with 5M, 28M, 123M, or 209M parameters. The family supports up to 32,768 historical steps and forecasts up to 1,024 steps per forward pass.
Enterprise customization
Through Mistral Forge, organizations can fine-tune on proprietary data, extend the architecture, and control training-data mixtures. The customization pipeline can run on client hardware, including air-gapped environments.
Who is it for?
Demand analyst planning inventory
Compare demand forecasts that incorporate historical sales and known promotional schedules.
Infrastructure team planning capacity
Forecast service-load metrics and compare central predictions with wider planning ranges.
Industrial data team analyzing sensors
Forecast related equipment measurements using historical readings and planned operating schedules.
How to use Timestral
- Install the official repository package in Python 3.10 or newer with PyTorch 2.4 or newer.
- Choose a checkpoint and load it through TimestralForecaster on CPU or GPU.
- Supply historical values as NumPy arrays or PyTorch tensors, adding covariates when relevant.
- Request a forecast horizon and retrieve the median and selected quantiles for evaluation.
Timestral pricing
Free
Downloadable weights are free for eligible users under the Modified MIT conditions. Organizations above the license's monthly revenue threshold need a separate commercial license from Mistral. Enterprise customization through Forge is available by inquiry.
Check current pricing ↗What to know before you use it
- Across all four checkpoints, the Modified MIT weight license excludes users whose company or employer exceeded USD 20 million, or its currency equivalent, in global consolidated revenue during the preceding month. Those users need a separate commercial license.
- History is limited to 32,768 steps. Horizons beyond 1,024 steps are generated block by block, feeding each block's predicted medians into the next context.
- Covariates are conditioning inputs and are not forecast themselves. Known-future covariates must provide values over both the historical context and requested forecast horizon.
Frequently asked questions
Do I need a GPU?
No. The standalone inference library runs on CPU as well as GPU through PyTorch.
Can forecasts run without network access?
Yes, once the required checkpoint files are local. The forecaster can load a local directory or use cached files in offline mode.
Do the code and weights share a license?
No. The repository code uses Apache 2.0, while the separately downloaded model weights use Modified MIT terms, including the company-revenue restriction.