Mistral says its first time-series foundation model family, Timestral, puts a 209M-parameter model first on GIFT-Eval, TIME and fev-bench at once, beating larger prior leaders with about two-thirds the parameters.
It is a small, CPU-friendly forecaster in a market of billion-parameter giants.
The company released four sizes at 5M, 28M, 123M and 209M parameters, the company said, all sharing one architecture with 32,768-step context and probabilistic outputs over 21 quantiles. Timestral handles multivariate targets and both past-only and known-future covariates, including categorical inputs, and Mistral said the 5M model forecasts one day ahead from three weeks of 15-minute history in 10 ms on an 8-thread server CPU while the 209M model takes 140 ms. The 28M model is five times faster than the 209M, and the 123M comes within 0.2% of the top score at three quarters of the latency, according to the company.
That Pareto claim is where context matters. Mistral said the 28M Timestral beats Chronos-2, roughly four times larger, on GIFT-Eval and TIME, and the 209M beats TimesFM-3 and Falcon-2.0 with fewer parameters. All scores are zero-shot from a single checkpoint without ensembling, the company said, with submissions made to the GIFT-Eval and TIME leaderboards. Results lean on synthetic-first pre-training at 8,192 context then post-training on real plus synthetic data at 4x longer context, plus decoupled horizon forecasting and covariate-aware attention.
The numbers leave open questions. Benchmark leadership does not equal production value, and Mistral has not yet published the technical report or independent leaderboard confirmations as of the October post. Pricing, license limits under the Modified MIT terms, and drift performance after fine-tuning via Mistral Forge for on-prem or air-gapped deployment are undisclosed. What would change the read is third-party replication on private enterprise data and latency and cost at scale, not just zero-shot skill scores.
