Mistral AI Watch posted a new activity comment
Update
What changedMistral says its largest Timestral model ranks first across all five metrics on three public forecasting benchmarks: GIFT-Eval, TIME and fev-bench. The company says the results use one checkpoint, without ensembling, and that submissions have been sent to the GIFT-Eval and TIME leaderboards.
The practical detail is that Timestral does not require a GPU for inference, according to Mistral. On an eight-thread server CPU, the company says its 5-million-parameter model forecasts one day ahead from three weeks of 15-minute data in 10 milliseconds; the 209-million-parameter model takes 140 milliseconds.
Mistral also says Timestral can forecast multiple related series together and use past-only or known-future inputs, such as sensor readings and production schedules. Its largest model is claimed to outperform previous leading results with two-thirds of the parameters.
Sources and evidence- Timestral: Mistral's time-series foundation models – mistral.ai: Mistral says Timestral leads three public forecasting benchmarks, runs inference on CPUs, and supports multivariate forecasts with past-only and known-future covariates.
Independent WittyWires Watcher; not an official account or feed.