Mistral's Timestral forecasts time series on a CPU in 10 to 140 milliseconds
Four sizes, from 5 million to 209 million parameters, all sharing one architecture and one training recipe.
Mistral released Timestral, its first time-series foundation models. The 209M model tops three zero-shot benchmarks with a single checkpoint, and the 28M model reportedly beats Chronos-2, which is about four times larger.
On an 8-thread server CPU, the smallest takes roughly 10 ms per forecast, with no GPU needed.
What Timestral is
Mistral's announcement introduces Timestral as the company's first family of time-series foundation models. They produce probabilistic, multivariate forecasts and can use both past-only and known-future covariates. The four models (5M, 28M, 123M and 209M parameters) share one architecture and training recipe, so you trade accuracy for speed by choosing a size.
The claims
Mistral says the 209M model ranks first among pre-trained, zero-shot models on GIFT-Eval, TIME and fev-bench, using a single checkpoint without ensembling, and beats the previous state of the art with about two-thirds of the parameters. The 28M model reportedly outperforms Chronos-2 on GIFT-Eval and TIME. These are Mistral's own numbers.
On an 8-thread server CPU, the 5M model takes about 10 ms per forecast and the 209M model about 140 ms.
Availability
The license is Modified MIT, with weights on Hugging Face and code on GitHub. Organizations can fine-tune through Mistral Forge, and deployment options include on-premises, VPC and air-gapped environments.
Two teams running baselines saw no gain, and the license has a $20M cap
One commenter runs about 400k per-series XGBoost models every 30 minutes at work:
"xgboost model per timeseries has never been beaten before." — the commenter running that fleet
"It did not beat our AutoARIMA baselines, they are pretty noisy series though." — a commenter who ran it on internal datasets
The license is narrower than "Modified MIT" suggests, according to a clause pasted after a software engineer who writes about compilers called it restrictive:
"You are not authorized to exercise any rights under this license if the global consolidated monthly revenue of your company (or that of your employer) exceeds $20 million (or its equivalent in another currency) for the preceding month." — a commenter who pasted the clause from the license
So check your employer's revenue against that clause first, then pit the 28M model against your current per-series baseline.
Get breakdowns like this every day
The top Hacker News stories, unpacked — by email, Telegram, or RSS.
Mac.apk runs Android apps as native Mac processes, with no emulator or VM
Of 66 apps the author tested, 52 are fully playable. A commenter says the newest release already moves saved games out of the app bundle.