Mistral's Timestral chart is the lab's ranking, not a second run
Mistral says its 209 million parameter Timestral model leads three public forecast charts. The scores are the lab's. This page does not rerun them.
The short version
Mistral says Timestral, in four sizes up to 209 million parameters, leads GIFT-Eval, TIME, and fev-bench among pretrained zero-shot models, from one checkpoint. Those rankings are the lab's submission, not a measurement this desk repeated.
The family the note actually released
Mistral's October 15, 2026 note introduces Timestral as its first family of time-series foundation models. The largest is 209 million parameters. The note says that model ranks first on GIFT-Eval, TIME, and fev-bench among pretrained, zero-shot models, across the metrics those boards use, from one checkpoint and without ensembling. It also says the results have been submitted to the GIFT-Eval and TIME leaderboards. A submission is the lab's packet. It is not a referee this desk convened. The document named in the sources is the record this page will keep.weights you can fetch.
Four sizes, one recipe
The note lists 5 million, 28 million, 123 million, and 209 million parameters. It says the larger models improve forecast performance and the smaller ones are faster and cheaper, and that all four share the architecture and the training recipe. It says each size sits on the front of accuracy versus size on the charts it shows. That sentence is a comparison the lab drew. A reader who wants the plotted points is in the note, not in a second lab's table.the artificial intelligence desk.
What the card is willing to bound
The Hugging Face card for Timestral-209M says the model supports univariate and multivariate forecasts, past-only and known-future covariates, and probabilistic predictions across 21 quantiles. It says the history window goes to 32,768 steps and the forecast to 1,024 steps. Those are interface bounds. They do not say how wrong a forecast is on a series the card does not name. Asked directly, How many sizes does the note list? Four: 5 million, 28 million, 123 million, and 209 million parameters. The note says they share an architecture and that each size sits on its accuracy-versus-size comparison.an evaluation a second lab can run.
What an October 18 retelling does not add
MarkTechPost's October 18 item is on the wire as an account of the release. It is a writeup of Mistral's note, not a new leaderboard. This page will not treat a retelling as confirmation that the rank survived a second implementation. If a public board later shows a different order, the note's sentence and the board's sentence can both be dated. Asked directly, What does the model card add about the window? The Hugging Face card for Timestral-209M says the model accepts up to 32,768 historical steps, forecasts up to 1,024 steps, and emits 21 quantiles. Those bounds are the card's.Futureweb.
Classical tools stay a different object
The note says ARIMA and linear regression are usually fit to one series, and that models such as PatchTST are often trained from scratch on each dataset. Timestral is described as pretrained, so a zero-shot forecast is the claim. A zero-shot rank does not retire a classical fit on a series someone already understands. The dependency is which document holds the score. Asked directly, Is MarkTechPost a second laboratory? No. The October 18 item retells the release. It is not a rerun of GIFT-Eval, TIME, or fev-bench.what on-device inference keeps.
What this page will not promote
It will not call the 209 million parameter model a general language model. It will not convert a quantile forecast into a single number the card did not choose. It will not invent a CPU timing the note's opening does not need for the ranking sentence. The chart is the lab's. The date is October 15 for the note and October 18 for the retelling. The document named in the sources is the record this page will keep. The document named in the sources is the record this page will keep.
Where a forecast still has to land
Time series, the note says, shows up as sensor telemetry, observability, capacity, energy, supply, and risk. A foundation model that transfers patterns is the product claim. A plant that needs a number for one asset still has to decide whether the public chart is the right test. This page will re-check if the named leaderboards publish a row that contradicts the note's first-place sentence. Asked directly, Does a first-place chart mean the forecast is a measurement of a particular plant? No. The note describes zero-shot forecasts on public benchmarks. A rank on those charts is not a capacity figure for one factory, one grid, or one account.
Covariates are inputs, not a guarantee
The note says the models support past-only and known-future covariates. A known-future covariate is a value the forecast is allowed to see, such as a schedule the operator already has. It is not a promise that the future arrived early. The card repeats the covariate support. Neither page prints a worked example this brief will copy. Asked directly, Who published the Timestral rankings? Mistral. The October 15 note says the 209 million parameter model ranks first on GIFT-Eval, TIME, and fev-bench among pretrained zero-shot entries, from a single checkpoint, without ensembling.
What stays on the record
October 15, four sizes, a 209 million parameter model, three named public benchmarks, a single-checkpoint claim, 21 quantiles, and a history and horizon the card bounds. October 18 adds a retelling. The rank stays labeled as Mistral's. The short answer this page will stand on is this: Mistral says Timestral, in four sizes up to 209 million parameters, leads GIFT-Eval, TIME, and fev-bench among pretrained zero-shot models, from one checkpoint. Those rankings are the lab's submission, not a measurement this desk repeated.
Sources
The reports this brief is filing. Futureweb did not republish them.
Questions
Who published the Timestral rankings?
Mistral. The October 15 note says the 209 million parameter model ranks first on GIFT-Eval, TIME, and fev-bench among pretrained zero-shot entries, from a single checkpoint, without ensembling.
How many sizes does the note list?
Four: 5 million, 28 million, 123 million, and 209 million parameters. The note says they share an architecture and that each size sits on its accuracy-versus-size comparison.
What does the model card add about the window?
The Hugging Face card for Timestral-209M says the model accepts up to 32,768 historical steps, forecasts up to 1,024 steps, and emits 21 quantiles. Those bounds are the card's.
Is MarkTechPost a second laboratory?
No. The October 18 item retells the release. It is not a rerun of GIFT-Eval, TIME, or fev-bench.
Does a first-place chart mean the forecast is a measurement of a particular plant?
No. The note describes zero-shot forecasts on public benchmarks. A rank on those charts is not a capacity figure for one factory, one grid, or one account.