Mistral releases Timestral, four open-weight time-series forecasting models from 5M to 209M parameters
Mistral's first time-series foundation models run on a CPU and, the company says, top the GIFT-Eval, TIME and fev-bench leaderboards, but the weights carry a revenue-limited licence.
· 6 min read
Written by AI from the sources linked below. How we use AI
On this page
Mistral AI has released Timestral, its first family of time-series foundation models: four forecasting models from 5 million to 209 million parameters that run on an ordinary CPU and, the company says, rank first among pretrained zero-shot models on the three main public forecasting benchmarks. The weights are on Hugging Face and the inference code is on GitHub. For anyone who forecasts demand, energy load, server metrics or sensor data, it is a small, downloadable model rather than a hosted API, which makes it easy to test. The benchmark claims come from Mistral and its model cards, and a technical report has not yet been published.
Key points
- Timestral comes in four sizes: 5M, 28M, 123M and 209M parameters, announced by Mistral on 15 October 2026.
- Mistral says the 209M model ranks first on GIFT-Eval, TIME and fev-bench among pretrained, zero-shot models, using a single checkpoint without ensembling.
- On GIFT-Eval the 209M model's MASE is 0.658 against 0.661 for t0-gamma and 0.667 for TimesFM-3, a narrow margin, according to its Hugging Face model card.
- On a standard 8-thread server CPU, Mistral reports 10 ms (5M) to 140 ms (209M) to forecast one day ahead from three weeks of 15-minute history.
- The code is Apache 2.0; the weights use a "Modified MIT" licence that bars use by companies with more than $20 million in monthly revenue unless they get a commercial licence.
What happened
Mistral published its announcement, "Introducing Timestral", on 15 October 2026. The model repositories on Hugging Face were created on 14 October, and the company published a GitHub repository, mistralai/timestral, with standalone inference code that needs only PyTorch, NumPy, safetensors and huggingface_hub.
A time-series foundation model is trained on a large, varied collection of sequences (trends, seasonal cycles, noise) so that it can forecast a series it has never seen without any training on that series. That is called zero-shot forecasting. Timestral produces probabilistic forecasts: instead of one number per future step, it outputs 21 quantiles (from 0.01 to 0.99), so a user can see how wide the uncertainty is. It handles several related series at once (multivariate), and accepts extra inputs called covariates, both past-only ones and ones known in advance, such as a holiday calendar, a production schedule or a weather forecast. Categorical covariates, for example the day of the week or a machine's state, are accepted directly.
According to the model card, the 209M model is a patch-based transformer with 24 layers and an embedding dimension of 768. It reads up to 32,768 historical time steps and forecasts up to 1,024 steps in one forward pass; longer horizons are produced block by block.
Background
Classical tools such as ARIMA are fitted to one series at a time. Deep-learning models such as PatchTST learn across many related series but are usually trained from scratch for each dataset. Foundation models for time series apply the pretraining idea from language models: train once on a broad corpus, then use the model on new data directly or fine-tune it. Mistral's announcement names TimesFM-3, Toto 2.0, Chronos-2 and Falcon-2.0 as the models it compares against, which shows how crowded the field has become.
Three public leaderboards act as the main scoreboards: GIFT-Eval, TIME and fev-bench. Mistral says it submitted results to the GIFT-Eval and TIME leaderboards. Its GitHub README states that at launch the 209M model ranked first among pretrained and zero-shot models on GIFT-Eval, and first across 98 tasks on TIME and 100 tasks on fev-bench.
The numbers
The model card compares the 209M model with four competitors. Lower MASE (mean absolute scaled error) is better.
The margins are small. The 209M model beats the second-best entry on this table by 0.003 MASE on GIFT-Eval, and on the TIME benchmark the card lists 0.626 against 0.637 for t0-gamma. On fev-bench, where higher is better, the SQL skill score is 50.39 against 49.42 for t0-gamma. Those figures are the company's own, taken from its model card, and the leaderboards are the place to confirm them.
The more interesting result is the trade-off between size and accuracy. The family table shows how little the largest sizes add:
| MASE | |
|---|---|
| 5M | 0.705 |
| 28M | 0.678 |
| 123M | 0.659 |
| 209M | 0.658 |
Source: Timestral model cards on Hugging Face
The 123M model is within a hair of the 209M model, and Mistral says it comes within 0.2% of the larger model's score at three quarters of its latency. The company also says the 28M model beats Chronos-2, which it describes as about four times larger, on GIFT-Eval and TIME, and that the 28M model is five times faster than the 209M one.
Speed is the second selling point. Mistral reports that on a standard server CPU with 8 threads, forecasting one day ahead from three weeks of 15-minute history takes 10 ms with the 5M model and 140 ms with the 209M model. Independent commentator Laurence Moroney, who wrote a walkthrough on 16 October, quotes the README's weight sizes: 21 MB (fp32) for the smallest model and 837 MB for the largest.
How it works
Mistral lists four ideas, and says a technical report will describe each in detail:
- Decoupled Horizon Forecasting. Earlier causal decoders mask out the values to be forecast, so they look like missing data and are also lost as context for later forecasts. Mistral's approach keeps the full history in the sequence and uses learned forecast tokens that read the history before them, so every observed value serves as both target and context.
- Covariate-aware attention. Each input series gets a role: target, past-only covariate or known-future covariate. Information flows from covariates to targets only, so a holiday calendar can help without leaking a target's own future.
- Synthetic-first data. Most training data was synthetic: generated univariate processes with trends, seasonality and noise, combined into multivariate series with nonlinear dependencies.
- Two-stage training. Pretraining on purely synthetic data with a context of 8,192 steps, then post-training on a mix of real and synthetic data with a context four times longer. The loss is a pinball loss over 21 quantiles.
Getting started takes a few lines of Python: install from the GitHub repository, load TimestralForecaster.from_pretrained("mistralai/Timestral-209M"), pass an array of shape (series, variates, time) and a prediction length, and read back the quantiles and the median.
What it means
For practitioners, the practical change is that a strong forecasting baseline can now be a pip install away and run without a GPU. A small team can try the 28M or 123M model on its own metrics in an afternoon and compare it against whatever it uses now. Because the output is a set of quantiles, alert thresholds can be set from the forecast's own uncertainty rather than from a fixed band.
For the model makers, the headline is the rate of leapfrogging: the margins between the top entries on GIFT-Eval are in the third decimal place, so the ranking may change with the next release from a competitor.
Mistral also positions Timestral as a component in larger systems: a forecaster that flags an anomaly on an industrial pump, with a language-model agent then looking up similar past failures and drafting a diagnosis. The company describes this as an example of what could be built, not as a shipped product.
What we don't know yet
- Independent verification. The rankings come from Mistral's announcement, README and model cards. The technical report "will follow", and we have not checked the live leaderboards ourselves.
- Training data. Mistral says most of the data was synthetic but has not published the real-data mix.
- Licence in practice. The weights' licence bars use by any company whose global consolidated monthly revenue exceeds $20 million (or the equivalent), and says the restriction applies to derivatives too; larger firms must ask Mistral for a commercial licence, which it may grant at its discretion. Hugging Face lists the licence as "other". Read the LICENSE file before building on it.
- Real-world accuracy. Public benchmarks may not match a given company's data; Mistral itself says production data can behave differently and recommends adaptation through its Mistral Forge product.
What to watch next
Watch for Mistral's technical report, for independent leaderboard entries and re-runs, and for responses from the makers of TimesFM, Chronos and Toto. Also watch whether the licence terms limit adoption among larger companies, and how many users fine-tune the smaller models on their own data.
Sources
Written by
The official account of the people who run audiotale.net: announcements, conversation starters and the newsroom's drafts. Posts written with AI say so. Every other account here is a real person.
Read next
Intel GPU D-series: up to 480GB LPDDR5X, 1.5TB/s and 350W for AI inference, now sampling
Intel named its Crescent Island inference GPU at OCP 2026; the 480GB figure is a ceiling, and Intel's own card reportedly carries 160GB.
audiotale.net team5 min read
Anthropic commits 5% of its shares to a new Public Benefit Fund
Anthropic's new fund starts with 5% of its shares, can grow by 2% for each 50% rise in company value, and will not be allowed to lobby or fund politics.
audiotale.net team6 min read
OpenCode breach: provider API keys saved in Console may have been stolen, 1.4% of users hit
OpenCode maker Anomaly Innovations says an attacker reached a legacy system holding third-party OpenAI and Anthropic keys; users should rotate any key saved in a Console workspace.
audiotale.net team6 min read