Four models, none of them large
Mistral has released Timestral, its first family of time-series foundation models, built for probabilistic multivariate forecasting rather than text. The announcement, dated 15 October, covers four checkpoints that share one architecture and one training recipe: 5M, 28M, 123M and 209M parameters. The Hugging Face collection lists them as Timestral-5M through Timestral-209M, with the smallest at 5.21 million parameters.
The practical claim is that none of them needs an accelerator. Mistral says that forecasting one day ahead from three weeks of history at 15-minute resolution takes about 10 milliseconds on the 5M model and about 140 milliseconds on the 209M, running on eight CPU threads. No GPU is required at inference.
The context window is 32,768 time steps. Inputs can include multivariate targets, past-only covariates and known-future covariates, with categorical covariates supported; the output is a probabilistic forecast across 21 quantiles rather than a single number.
The benchmark claims are the company’s own
Mistral reports that the 209M model ranks first on three public forecasting benchmarks — GIFT-Eval, TIME and fev-bench — and says it does so “with only two-thirds of the model parameters” of the previous state of the art. Among pre-trained zero-shot models it claims first place on all five metrics, using a single checkpoint with no ensembling, and places in the top ten overall of 83 entries on GIFT-Eval, 35 on TIME and 37 on fev-bench.
Two specific comparisons stand out. The 28M model, Mistral says, beats Chronos-2 on GIFT-Eval and TIME despite Chronos-2 being roughly four times larger. The 209M beats TimesFM-3 and Falcon-2.0 with fewer parameters. The 123M lands within 0.2% of the 209M score at about three-quarters of its latency.
All of these are vendor-reported figures, published without an accompanying technical report — Mistral says that is still to come — and without an independent evaluation. The leaderboards named are public, so the rankings are checkable in principle; the parameter-efficiency framing around them is Mistral’s.
What is actually shipping
The weights are released under a modified MIT licence, available through a GitHub repository and a Hugging Face collection, with domain adaptation handled through Mistral Forge. The company lists on-premises, VPC and air-gapped deployment, which together with the CPU-only inference story points at the intended buyer: an operations team forecasting inside its own network rather than calling an API.
The use cases Mistral names are capacity planning, energy, supply chains and sensor telemetry. Those are domains where the data rarely leaves the building and where a 28-million-parameter model that runs on existing hardware is a materially different proposition from a hosted frontier model.
Why the size matters more than the ranking
Time-series foundation models have been converging on the same bet for two years: that pre-training on enough series, largely synthetic, produces a forecaster that works zero-shot on data it has never seen. Mistral credits synthetic pre-training, covariate-aware attention, a technique it calls Decoupled Horizon Forecasting, and a two-stage training run that starts at 8,192 context and then extends it fourfold.
If the efficiency claims hold up independently, the interesting number is not the leaderboard position but the parameter count attached to it. A forecaster small enough to sit beside the system it is forecasting changes where this kind of model can be deployed, and by whom.
The things to watch are the technical report, which has not appeared, and the first independent runs on GIFT-Eval and fev-bench by someone other than Mistral. Until then the rankings should be read as what they are: the vendor’s own evaluation of its own release.
