Mistral releases Timestral, a family of four time-series foundation models (5M–209M parameters) claiming top ranks on GIFT-Eval, TIME and fev-bench

Mistral AI has released Timestral, its first family of time-series forecasting models, in four sizes: 5M, 28M, 123M and 209M parameters. They produce probabilistic, multivariate zero-shot forecasts (no training on the user's data needed) and accept extra inputs such as known-future and categorical variables. Mistral says they run on a CPU, from about 10 ms (smallest) to 140 ms (largest) for a one-day-ahead forecast from three weeks of 15-minute data on an 8-thread server. Results have been submitted to the GIFT-Eval and TIME leaderboards, fine-tuning is offered through Mistral Forge, and a technical report is promised but not yet published.

Teams that forecast demand, energy load, sensor readings or operational metrics now have another small, cheap-to-run option alongside Chronos-2 and similar models. CPU-only inference at these speeds, if it holds up, lowers the cost of embedding forecasts in monitoring or anomaly-detection pipelines. It also shows Mistral moving beyond language models into a narrow enterprise niche that connects to its Forge fine-tuning offer. The practical impact is limited until independent users test it on their own data and the licence and weight availability are clear.

The "first on all three benchmarks" and "every size on the Pareto front" claims are Mistral's own reporting, with no technical report yet and no linked primary sources. Leaderboard rank gaps in this field are often small and can reflect benchmark overlap with training data, so they say little about performance on a given company's data. The agentic anomaly-detection and root-cause framing is aspiration, not demonstrated capability.

1. 01Publication of the technical report, including training data sources, any check for benchmark leakage, and full results on GIFT-Eval, TIME and fev-bench.
2. 02Whether the leaderboard entries appear publicly and are held or reproduced by independent parties, especially the claim that the 28M model beats Chronos-2.
3. 03Licence and weight availability (open weights or API only), plus pricing for Mistral Forge fine-tuning, and any third-party reports on real-world data within the next one to two months.