跳到正文
研踪PaperTrail
← 研踪 PaperTrail

arXiv:2610.37585 · v1

Timestral

Adrien PetraliaAlexandre GavaudanArtjom JoosenEtienne GoffinetRichard KurleYassine El OuahidiYassir BendouAbdelaziz BounharAbhijeet SomaniAditi KabraAdrian ValenteAdrien SadéAhmed DjermaniAlan JeffaresAlbert JiangAlberto BiettiAleksandr TimashovAlexandre CahillAlexandre LavalAlexandre SablayrollesAlexandros Sanchez VassopoulosAmélie HéliouAmos YouAndré JonassonAndreas FürstAndrew BaiAndrew EhrenbergAndrew ZhaoAndy LoAngele LenglemetzAnmol AgarwalAntonia CalviArata SuzukiArjun MajumdarArthur FournierAshwani KumarAugustin LombardAvinash SooriyarachchiAylin Guliz AkkusAysenur KaradumanBaptiste BoutBaptiste RozièreBarna PásztorBaudouin De MonicaultBenjamin HolzschuhBenjamin TibiBenno SteinBernhard StadlbauerBhargavi BhaskarBinbing HouBłażej OsińskiBoris van BreugelCamille Le ScaoChaoran YuCharlotte CronjägerChen-Yo SunChris BamfordChristian WallenweinChristophe RenaudinClémence LanfranchiColin BonattiCorentin BarreauCorentin SautierCristiana-Diana DiaconuCyprien CourtotDaniel LopesDaniel MarczakDarius DabertDiego de las CasasDominik NussDylan RubiniDzmitry SoupelElizaveta DemyanenkoElliot Chane-SaneEmilien FugierEmmanuel GottlobErik AasEtienne DuchesneFabian PaischerFabian SchlagerFaruk AhmedFederico BaldassarreFélix RosseeuwFilip SzatkowskiFilippo VicentiniFlorian WiesnerGabrielle BerradaGaëtan EcrepontGaspard BlanchetGaspard Donada-VidalGaurav MahajanGauthier GuinetGenevieve HayesGeorgii NovikovGeorgios ChristoglouGiada PistilliGonzague de CarpentierGraeme NailGuillaume BretonGuillaume KunschGuillaume LampleGuillaume MartinGuillaume RailleGuillaume RichardGuillem SimeonGunjan DhanukaGunshi GuptaHamza Senhaji RhaziHan ZhouHédi HadijiHimanshu JajuHope McGovernHubert de ParsevalHugo CisnerosHugo GermainHugo ThimonierHugo VenturiniHyeonju LeeHyunwoo OhIliass LasriIlyas OulkaddaJacques SunJan LudziejewskiJason RuteJean de BecdelièvreJean QuentinJean-Malo DelignonJessica ChudnovskyJianchao LiJie ZhangJin Bin OuJoachim StudniaJoep BarmentloJohannes BrandstetterJohn HarvillJonas AmarJonas SchweizerJonathan RobertsJoséphine DelasJosselin SomervilleJules DepersinJulien DenizeJulien TauranJulien VerlaguetKartik KhandelwalKhyathi Raghavi ChanduKilian TepKonstantin MishchenkoKris SzczepaniakKush JainLarissa LaichLaura CalemLaurence AitchisonLaurent CallotLaurent FainsinLéo CotteleerLéonard BlierLoïc ChadoutaudLoïck ChambonLorenzo ProiettiLouis FauconLouis MartinLouis SerranoLu LiLuc BoudierLucas Fernandes MartinsLucas MebilleLucile SaulnierLudovic Ho FuhLuis MonteroŁukasz BorchmannMaarten BuylMalik Hassan RazaManon ChossegrosMarcin MożejkoMargaret JenningsMarkus HennerbichlerMartin AlexandreMathieu PoiréeMathieu SchmittMathilde GuillauminMathis FelardosMatthieu AndréMatthieu DinotMatthieu FuteralMaurits BleekerMauro ComiMax MynterMaxim BermanMaxime DarrinMaxime LangelierMaxime LouisMaximilian AugustinMaximilian MüllerMehdi BouazizMelina Jingting LaimonMert UnsalMia ChiquierMichael PilcerMichał PietruszkaMichał ZającMickaël SeznecMikhail BiriuchinskiiMinh-Quang PhamMinwoo KangMircea LicaMorgane RivièreNamit KatariyaNatasha Ørregaard KlingenbrunnNathan BoutNathan GrinsztajnNathan SimpsonNeeraj AggarwalNeil AshtonNeil MallinarNick BührerNicolas MoreauNicolò LucchesiOla MysiakOliver LeichtOlivier BousquetOlivier KochParag JainPatricia WangPatrick BliesPatrick von PlatenPatryk SafferPaul JacobPaul TeiletchePaul WamberguePaula KurylowiczPavan Kumar ReddyPavel KuksaPhilippe FormontPhilippe PinelPhilomène ChagniotPierre GlandonPierre StockPierre-André SavallePiotr MiłośPoojan ShahPrateek GuptaPravesh AgrawalPui Ki ChanQuentin DesreumauxQuentin SelleQuentin TorrobaQuentin WargnierQuercus HernandezRam RamrakhyaRandall IsenhourRanjit ParvaRaphaël BaloukaRaphaël RiallandRaul Perez PelaezReinhard SonnleitnerRemi DelacourtRishi ShahRob RomijndersRohan ShahRohin AroraRomain SauvestreRoman SoletskyiRonan RiochetRosalie MillnerRupert MenneerSagar VazeSamuel BarrySamuel BelkadiSamuel HumeauSamuele PapaSanchit GandhiSandeep CranganoreSandeep SubramanianSarthak MittalSaskia AdaimeScott MatthewsSean ChaSebastian KaltenbachSeungjae JungShashwat VermaSherif WalyShrimai PrabhumoyeSiddhant WaghjaleSiddharth GandhiSimon LepageSimon SorgSoham GhoshSophie MarbachSrijan MishraStanislas LangeSteve HongSumukh AithalSzymon AntoniakTarun Kumar VanganiTeddy TurquetTeven Le ScaoThéo CachetThibaut LavrilThomas ChabalThomas CosteThomas DefardThomas FoubertThomas LouxThomas RobertThomas WangTianyu ZhangTim LawsonTimothée LacroixTimothy WallisTobias KronlachnerTom BewleyTom EdwardsTomas HodanTriantafyllos AfourasTuhin DasTyler WangUdit SaxenaUlrick BléUmar JamilUmberto TomasiniValentin MacéVan PhungVenkata ReddyVibha MastiVictor JouaultVictor LetzelterVictor PaltzVictor PoucheretVincent MaladièreVincent PfisterVirgile RichardVladislav BataevWassim BouazizWen Ding LiWilliam FlanaganWilliam HavardWilliam MarshallWitold JędrzejewskiXinghui LiXingran GuoXinyu YangXuanyu ZhangYann DrezeYann ViegasYannic NeuhausYassine TurkiYessin MoakherYi HeYihan WangYimu PanYuying ZhuYves Martin des TailladesZaccharie RamziZhenlin XuZsófia Csákány
首次提交 2026-10-28 01:25
版本更新 2026-10-28 01:25
雷达观测 2026-10-29 00:34
arXiv ↗PDF ↗

摘要

We introduce Timestral, a family of four time series foundation models for probabilistic forecasting, with 5M, 28M, 123M, and 209M parameters. Timestral sets a new state of the art on GIFT-Eval, TIME, and fev-bench, where Timestral-209M and Timestral-123M take the top two places on every metric. We achieve this with three ingredients: Decoupled Horizon Forecasting (DHF), which separates prediction tokens from the causally encoded context, so that longer horizons are trained without masking context and each observed value is both a forecast target and context for later forecasts; a covariate framework that natively supports known-future covariates within a causal architecture; and a two-stage pipeline that pretrains exclusively on a diverse, large-scale synthetic corpus before post-training on a mix of real and synthetic data with extended context. This report describes the model design and its benchmark results. Weights for all four models are available on Hugging Face and inference code on GitHub.

作者备注:Webpage: https://mistral.ai/news/timestral ; Weights: https://huggingface.co/collections/mistralai/timestral ; Code: https://github.com/mistralai/timestral

论文发现出处

这些页面提供论文线索,不代表同行评审或质量认证。

AI 论文解读

基于摘要 · 非全文审阅

此版本尚未生成 AI 解读。解读由后台任务生成并按论文版本保存,登录后共享已有结果。

版本差异

目前只观测到一个版本,无法比较变化。之后收录新版本时,会在这里对照标题、作者、分类和摘要。

观测到的版本

相邻记录可能跳过中间版本;首次发现 v2 及以上也会标为版本更新。

v1 · 2026-10-28 01:25 · 最新观测版本

Timestral

We introduce Timestral, a family of four time series foundation models for probabilistic forecasting, with 5M, 28M, 123M, and 209M parameters. Timestral sets a new state of the art on GIFT-Eval, TIME, and fev-bench, where Timestral-209M and Timestral-123M take the top two places on every metric. We achieve this with three ingredients: Decoupled Horizon Forecasting (DHF), which separates prediction tokens from the causally encoded context, so that longer horizons are trained without masking context and each observed value is both a forecast target and context for later forecasts; a covariate framework that natively supports known-future covariates within a causal architecture; and a two-stage pipeline that pretrains exclusively on a diverse, large-scale synthetic corpus before post-training on a mix of real and synthetic data with extended context. This report describes the model design and its benchmark results. Weights for all four models are available on Hugging Face and inference code on GitHub.

查看此版本解读 →查看此版本原文 ↗