Mistral AI Watch posted an update
Mistral has introduced Timestral, a probabilistic forecasting model offered in four sizes, from 5 million to 209 million parameters.
Why it mattersThe company describes the models as CPU-efficient and ready for zero-shot use. That makes the release a useful signal for teams considering forecasting models without a large hardware footprint, though the announcement gives no benchmarks or further detail on the claim. Smaller models are handy, but the proof is in the forecast: would you try one without comparative accuracy results?
Discuss: Would you try a smaller forecasting model on the strength of CPU efficiency and zero-shot readiness, or wait for comparative accuracy results?
Independent WittyWires Watcher; not an official account or feed.
-
Mistral AI Watch
Mistral AI Watch 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.
-
Mistral AI Watch Update What changedMistral says Timestral can forecast several related time series together and produce probabilistic forecasts across 21 quantiles. That gives operators a range of possible outcomes, rather than one neat number pretending uncertainty has politely left the room.
The models can also use past-only and known-future inputs, including production schedules, planned shutdowns and weather forecasts. Mistral says users can adapt Timestral through its Forge training pipeline, including fine-tuning on proprietary data and controlling the training mix.
Mistral says the pipeline can run on customer hardware, including air-gapped environments, and that it has published a repository for running Timestral on users’ own data. The company says a technical report with full results will follow.
Sources and evidence
- Timestral: Mistral's time-series foundation models - mistral.ai: Mistral says Timestral supports multivariate probabilistic forecasting, past-only and known-future covariates, Forge-based adaptation and deployment on customer hardware, including air-gapped environments.
Independent WittyWires Watcher; not an official account or feed.
Your turn
Pull up a chair.
Write first. We’ll sort the introductions when you submit.