Mistral releases Timestral time-series foundation models
Original title: Timestral: Mistral release time-series foundation models
Mistral says its 209M model ranks first on GIFT-Eval, TIME and fev-bench while using about two-thirds the parameters of the prior leader.
Mistral released Timestral, a four-size family of time-series foundation models from 5M to 209M parameters, for probabilistic multivariate forecasts with past and known-future covariates. The company says the 209M model ranks first among pretrained zero-shot entries on GIFT-Eval, TIME and fev-bench from a single checkpoint, and that every size sits on its accuracy-versus-size frontier. Inference is described as CPU-only, with Mistral Forge offered for domain fine-tuning. Results are attributed to Mistral and described as submitted to the GIFT-Eval and TIME leaderboards.
Unconfirmed: Benchmark ranks, parameter comparisons and latency figures are Mistral's own reported results.
- Sizes are 5M, 28M, 123M and 209M parameters, sharing one architecture.
- Mistral says the 28M model beats Chronos-2 on GIFT-Eval and TIME.
- On an 8-thread server CPU, one-day forecasts take 10 ms at 5M and 140 ms at 209M.
- The 123M model is within 0.2% of the 209M score at three quarters of its latency, Mistral says.
- Models accept categorical covariates and can run beside Mistral language models in agent workflows.
- Vendors
- Mistral
- Products
- Mistral Forge
- AI models
- Timestral
From the original article
6 min read at mistral.ai
Timestral, Mistral’s first family of time-series foundation models, delivers state-of-the-art probabilistic, multivariate forecasts with just 209M parameters, outperforming larger models on key benchmarks while supporting past and future covariates.
By groar · mistral.ai
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.