By asahifile 
Mistral has released Timestral, its first family of time-series foundation models — and the headline number is not the accuracy claim but the size. The largest model, at 209 million parameters, ranks first among pre-trained zero-shot models on the three public forecasting benchmarks the field uses most: GIFT-Eval, TIME and fev-bench. It runs on a CPU.
The family comes in four sizes — 5M, 28M, 123M and 209M parameters — sharing one architecture and one training recipe. On a standard eight-thread server CPU, Mistral reports, forecasting one day ahead from three weeks of 15-minute readings takes 10 milliseconds with the 5M model and 140 milliseconds with the 209M model. The fp32 weights range from 21 MB to 837 MB.
What a time-series foundation model actually does
A time series is a sequence of measurements over time: hourly electricity demand, a pump’s vibration readings, server request counts. Classical tools such as ARIMA fit one series at a time. Deep models such as PatchTST learn nonlinear patterns across related series, but are usually trained from scratch on each dataset, so what they learn does not carry to a new domain.
A time-series foundation model brings the pre-training trick to this data. Trained on a large, varied corpus, it learns patterns that transfer, and forecasts zero-shot on series it has never seen — including domains where history is scarce. Timestral accepts up to 32,768 past time steps, forecasts up to 1,024 steps in a single forward pass, and returns probabilistic output as 21 quantiles per step rather than a single line. It handles multivariate targets and both past-only and known-future covariates, including categorical ones such as day of the week or a machine’s state. Missing values go in as NaN.
The benchmark picture, and its limits
On GIFT-Eval, the 209M model records a MASE of 0.658, against 0.661 for t0-gamma, 0.667 for Google’s TimesFM-3 and 0.698 for Amazon’s Chronos-2, where lower is better. On fev-bench it scores a SQL skill score of 50.39, ahead of TimesFM-3 at 48.99. The margins at the top are thin: 0.003 separates Timestral from t0-gamma on GIFT-Eval.
These are Mistral’s own numbers, submitted to the GIFT-Eval and TIME leaderboards, and leaderboard placement can move as independent evaluations land. The more interesting comparison is the size-accuracy curve. A 209M model that is competitive with models several times larger, and that runs on CPU, changes the deployment arithmetic for anyone monitoring thousands of sensor streams.
Why CPU inference matters more than the leaderboard
Most enterprise forecasting happens next to the data: in a plant, on a gateway, inside a monitoring pipeline. Sending every sensor series to a GPU cluster is expensive and often impossible in regulated or air-gapped environments. Mistral lists on-premises, VPC and air-gapped deployment, and pairs the numerical model with its language models in agentic workflows — Timestral flags an anomaly on a pump, an agent compares the signal against embeddings of past failures, sub-agents pull maintenance records and shift logs, and a plant manager gets a diagnosis rather than an alert.
There is a licensing catch worth reading closely. The weights ship under a “Modified MIT” licence whose modification is a revenue cap: a company whose global monthly revenue exceeded $20 million in the preceding month is not authorised to use the model. Small teams can run it free; large enterprises are pushed toward Mistral Forge for fine-tuning and a commercial agreement. The technical report describes four contributions, including Decoupled Horizon Forecasting — keeping the full history in context instead of masking forecast targets, as contiguous patch masking does — and covariate-aware attention that lets information flow only from covariates to targets, preventing future leakage.
What to watch
- Independent leaderboard placement on GIFT-Eval and TIME once third-party evaluations are published.
- Whether Mistral ships a hosted Timestral API, and at what price, given that the weights are downloadable.
- Adoption in industrial monitoring and observability, where the CPU-only pitch is strongest.
- Whether the $20 million revenue cap becomes a template for open-weight commercial licensing.
- How practitioners handle the black-box problem — a 209M model that forecasts well but exposes no dynamics is hard to sign off in safety-critical operations.
Sources: Mistral AI; Timestral-209M model card, Hugging Face; MarkTechPost; Mistral’s GitHub repository; Hacker News discussion.