Mistral AI Releases Timestral: A 209M Forecasting Model Family That Outperforms Google’s TimesFM-3
MarkTechPost Asif Razzaq
Mistral released Timestral, a family of four forecasting models sized from 5M to 209M parameters. It runs on CPU, and the biggest model tops several zero-shot benchmarks.
Based on reporting by MarkTechPost, Asif Razzaq — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Mistral AI has released Timestral, its first family of time series foundation models. The lineup has four dense, zero-shot forecasting models, from 5M parameters at the small end to 209M at the top. The pitch is simple: better forecasts, less hardware drama.
The biggest model is the one making the noise. Mistral says Timestral-209M ranks first among pretrained zero-shot models on GIFT-Eval, TIME and fev-bench. It does that without needing a GPU at inference, which is the sort of detail that tends to matter once a model leaves the slide deck and lands on an actual server.
Timestral is built for probabilistic forecasting. It outputs 21 quantiles per time step, handles multivariate targets jointly, and can take past-only, known-future and categorical covariates. The model also supports a long context window of up to 32,768 time steps, and a single forward pass can forecast up to 1,024 steps. Longer horizons are handled block by block, with each block’s median fed back in as context.
Under the hood, Mistral describes Timestral as a causal, patch-based transformer with time attention and variate attention. The 209M version has 24 layers and an embedding dimension of 768, while each patch covers 32 time steps. The training setup is unusual too: Mistral says most of the data is synthetic, with trends, seasonality, noise and nonlinear cross-series dependencies, followed by a second stage on mixed data with four times longer context.
On the benchmarks listed in the model card, Timestral-209M posts a GIFT-Eval MASE of 0.658 and CRPS of 0.446, TIME MASE of 0.626 and CRPS of 0.522, and a fev-bench SQL skill score of 50.39. Mistral also says the 28M model beats Chronos-2 on GIFT-Eval and TIME, even though Chronos-2 is about four times larger. On a standard server CPU with 8 threads, Mistral reports the 5M model at 10 ms and the 209M model at 140 ms for a one-day-ahead forecast from three weeks of 15-minute history.
My take — AI-written commentary, not fact-checked reporting
This is the kind of release that makes the open-model crowd happy for a good reason: useful forecasting models that run on a normal CPU, with weights people can actually inspect. The awkward bit for the rest of the industry is that “bigger” keeps looking less like a strategy and more like a tax. Closed models can keep the marketing gloss; the boring, deployable stuff is starting to look pretty open already.
Read more about this at: MarkTechPost