SSyncDev

Timestral 1.0

SyncDev model directory · Profile updated Nov 17, 2026

Compare with another model

Timestral 1.0 is Mistral's open time-series forecasting family in four sizes (5M-209M), producing zero-shot probabilistic forecasts on CPU.

Key specifications

Specification score
15.0
Context window
32.77K
License
Modified MIT

The specification score summarizes recorded features and limits. It is not a benchmark result.

Pricing conditions: No API price stated; weights downloadable. Fine-tuning through Mistral Forge is a commercial service.

Inputs and outputs

Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.

Accepts

  • Text input

Produces

  • Text output

Capabilities

Open weights

What is Timestral 1.0?

Mistral had never shipped a forecasting model before October 2026. Timestral 1.0, released on 15 October, is its first family of time-series foundation models, and it arrives as four sizes rather than one: 5M, 28M, 123M and 209M parameters, all sharing an architecture and a training recipe.

That size spread is the interesting part. A 5M-parameter model can sit inside a service that already runs on a small VM. No GPU is required for any of them.

The four sizes

Size CPU latency (example task) Notes
5M About 17 ms Smallest; for very high-volume or embedded use
28M About 3.6x faster than 209M Mistral reports it beats Chronos-2, a model roughly four times larger, on GIFT-Eval and TIME
123M About three-quarters of 209M's latency Within 0.2% of the 209M score
209M About 113 ms Flagship

The latency figures come from one setup: an 8-thread CPU forecasting one day ahead from three weeks of 15-minute history. Your series length and covariates will change them.

What a forecast looks like

Each call returns probabilistic, multivariate forecasts as 21 quantiles, not a single line. That suits capacity planning and inventory, where the 90th percentile matters more than the median. Context runs to 32,768 time steps. Covariates come in three kinds: past-only, known-future (holidays, scheduled promotions) and categorical.

It works zero-shot. You can point it at a new series without training.

Publisher-reported results

For the 209M model, single checkpoint and no ensembling, Mistral reports GIFT-Eval GMean MASE of 0.6575 and WQL of 0.4460, TIME GMean CRPS of 0.5223, and an fev-bench skill score of 50.39%, and claims first place among pre-trained zero-shot models on all three suites. It also says the 209M edges out TimesFM-3 with fewer parameters. These are the publisher's numbers.

Licence and the open-weights question

The model table lists Timestral as open under a Modified MIT licence, and Mistral publishes a GitHub repository and a Hugging Face collection for it. Mistral's launch post never uses the words "open weights". In practice the checkpoints are downloadable, but read the modified licence terms before shipping in a commercial product, since "modified" is doing real work there.

Training was mostly synthetic: generated series with trends, seasonality and noise, plus a synthesiser for cross-series dependencies, followed by post-training on a real-and-synthetic mix at four times the original context. Fine-tuning on proprietary data is sold through Mistral Forge, which can deploy on-premises or air-gapped.

Backtesting before you switch

Hold out the last few months of your own data and roll the forecast origin forward week by week. Score pinball loss on the quantiles you act on, not just point error. Compare against your current baseline and against TimesFM 3.0. Start with the 28M size: if it is close enough, the latency and cost savings over 209M are large.

There is no stated maximum horizon. Test the horizons you need explicitly.

More in time-series forecasting, and Mistral's other models are on its provider page.

How to evaluate Timestral 1.0

Compare the model on the work you need to complete.

Tasks and inputs

Timestral 1.0 is listed for Time Series Forecasting. It accepts text and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.

Capacity and cost

Documented limits: 32.77K context. Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.

Choosing a deployment

The weights are downloadable under Modified MIT. Check runtime support and size the complete model and working context for your expected concurrency.

Frequently asked questions

Four: 5M, 28M, 123M and 209M parameters, sharing one architecture and training recipe.

These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.

allenai

ACE2.5-ERA5

ACE2.5-ERA5 is Ai2's open stochastic climate emulator trained on ERA5, producing calibrated atmospheric ensembles from days to centuries at low compute.

— context

allenai

HiRO-ACEv2

HiRO-ACEv2 is Ai2's open AI climate emulator that produces 3 km precipitation, winds and pressure from a 100 km run, over 300x cheaper than a numerical model.

— context

IBM

Granite TimeSeries PatchTST-FM-r2

Granite TimeSeries PatchTST-FM-r2 is IBM's open ~385M zero-shot forecasting model, dual-licensed Apache-2.0/OpenMDW, with 8,192-step context.

8.19K context

Google

TimesFM 3.0 PyTorch

TimesFM 3.0 is a numerical time-series forecasting checkpoint for PyTorch under a custom noncommercial licence.

— context