A 209M-parameter forecasting model that runs on CPUs and handles past and future covariates zero-shot gives teams a small, cheap option for probabilistic forecasting. benchmarkA standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.Full definition → rankings are self-reported and submitted to leaderboards.
Key takeaways · AI-distilled
Timestral needs no GPU for inferenceRunning a trained model to get answers — the phase where AI is actually used, as opposed to trained.Full definition →, Mistral says: on an 8-thread server CPU, a one-day forecast from three weeks of 15-minute data takes 10 ms on the 5M model and 140 ms on the 209M model.
The family comes in 5M, 28M, 123M and 209M sizes sharing one architecture. Mistral reports the 123M model lands within 0.2% of the 209M score at three quarters of its latency, and the 28M model runs five times faster than the 209M.
Covariate-aware attentionThe mechanism that lets a model weigh which earlier words matter for the word it's currently processing — the core operation of a transformer.Full definition → gives each input series a role (target, past-only or known-future covariate) and lets information flow only from covariates to targets, so calendars or weather forecasts can be used without leaking a target's own future.
Most training data was synthetic: Mistral pre-trained on purely synthetic series at context windowThe maximum amount of text a model can consider at once — its working memory for the current conversation or task.Full definition → 8,192, then post-trained on mixed real and synthetic data with a 4x longer context, using a pinball loss over 21 quantiles.
The model handles categorical covariates such as machine state directly, outputs 21 quantiles, ships under a modified MIT licence, and can be deployed on-prem, in a VPC or air-gapped. A GitHub repo is out; the technical report is still to come.
Terms in this piece · Glossary
benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
inference — Running a trained model to get answers — the phase where AI is actually used, as opposed to trained.
attention — The mechanism that lets a model weigh which earlier words matter for the word it's currently processing — the core operation of a transformer.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
