Best Tools · Models
Timestral
SummaryMistral's CPU-friendly time-series models that forecast from history with uncertainty bands, no per-series training.
Filed October 15, 2026. Price, as published by the maker: Code free (Apache-2.0). Weights Modified MIT — no rights if your company (or employer) monthly revenue exceeded $20M last month..
Open the maker’s site (opens in a new tab)
Timestral is Mistral's family of pre-trained time-series models (5M, 28M, 123M, 209M), announced October 15, 2026. You give a history of numbers and get a forecast with quantile bands. The inference code on GitHub is Apache-2.0; the weights on Hugging Face use a Modified MIT licence that grants no rights if your company's (or employer's) global consolidated monthly revenue exceeded $20 million in the preceding month.
How we tested. October 22, 2026, on a CPU-only Linux box (PyTorch 2.14.1, 4 threads). Package 1.0.0 from pip install git+https://github.com/mistralai/timestral.git (commit a534bdb). Data: daily English Wikipedia pageviews for "Large language model", Jan 1, 2024 to Oct 21, 2026 (1,025 days). We forecast 28 days ahead from 8 starting points against a "same as last week" repeat.
- Timestral-28M (111 MB): average error 270 views/day vs 345 for the repeat (21.5% lower), won 8 of 8, ~22 ms per forecast.
- Timestral-5M (21 MB): 270 vs 345 (21.6% lower), 8 of 8, ~8 ms.
- Timestral-123M (490 MB): 276 (19.9% lower), won only 7 of 8.
- The 10–90% band held the real number on 80–84% of days.
One series only — a sanity check, not a benchmark. We did not test the 209M model, covariates, fine-tuning or long horizons. Mistral's leaderboard claims are Mistral's.
Pro: Small enough to run on CPU; the 5M matched the 28M on our series.
Con: Licence revenue cap; pattern-only forecasts (no causal drivers); bigger was not better on our one series.