Overview
- Mistral announced the Timestral family in October 2026 with four sizes (5M, 28M, 123M, 209M) and submitted zero-shot leaderboard results claiming the 209M model ranks first on GIFT-Eval, TIME and fev-bench.
- Timestral uses a single architecture with a 32,768-step context, probabilistic outputs over 21 quantiles, support for multivariate targets and known-future covariates, and a training recipe Mistral describes as synthetic-first pretraining followed by real-plus-synthetic post-training at longer context.
- Mistral published CPU latency claims for an 8-thread server: 10 ms for the 5M model and 140 ms for the 209M, and the company framed the smaller models as offering a favorable accuracy-versus-latency trade-off for production use.
- Independent testing by a developer who ran the public weights confirmed point-forecast accuracy and reproduced faster CPU times with a thread-count caveat but found the model’s probabilistic intervals were badly miscalibrated on a trending series, raising questions about uncertainty estimates in real deployments.
- Key gaps remain: Mistral has not released a technical report or independent leaderboard confirmations, and it has not disclosed pricing, license limits under the Modified MIT terms, or how the models handle drift and fine-tuning in enterprise on-prem or air-gapped setups using Mistral Forge.