Mistral AI · Oct 15, 2026
Mistral AI announces Timestral, a compact family of time-series forecasting models
Markdown · JSON · Original source
Mistral AI describes Timestral as its first family of time-series models trained broadly on data to forecast patterns in sequences, such as sensor readings. The four models range from 5 million to 209 million learned numerical settings and produce forecasts across related data series that represent uncertainty. Mistral reports that its largest model ranked first on three public benchmarks, and says the models can run forecasts on a standard server CPU. The company says most training data was synthetic and that a technical report with full results will follow.
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
Mistral AI · Article
Source details
Original source date: Oct 15, 2026 (2026-10-15T12:00:00.000Z).
Date basis: Date supplied by the publisher’s feed.
Added to Escape Meridian: 2026-10-21T12:27:39.637Z. First observed: 2026-10-15T15:06:37.276Z. Retrieved: 2026-10-16T02:29:03.772Z.
Referenced by
Recorded citations and public file attachments. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
Benchmark rankings and performance figures are reported by Mistral in the supplied publisher page; the evidence does not independently verify them. Mistral says results were submitted to the GIFT-Eval and TIME leaderboards and that a technical report with full results will follow.
Open question
How well will Timestral’s reported benchmark performance transfer to real-world forecasting tasks?
Source-stated
Mistral describes Timestral as its first family of time-series foundation models.
Today, we’re releasing Timestral, our first family of time-series foundation models.
Source-stated
Mistral says the four-model family ranges from 5 million to 209 million parameters and supports probabilistic forecasts across multiple related series, with past and known-future covariates as inputs.
Timestral delivers probabilistic, multivariate forecasts and supports past-only and known-future covariates.
Source-stated
Mistral reports that its 209-million-parameter model ranked first on three public forecasting benchmarks; it says the results were submitted to the GIFT-Eval and TIME leaderboards.
Our largest model, at just 209M parameters, ranks first on all three of the field's main public benchmarks, GIFT-Eval, TIME, and fev-bench, outperforming the previous state of the art with only two-thirds of the model parameters.
Source-stated
Mistral says Timestral can run inference on a standard server CPU and gives example runtimes for forecasts one day ahead from three weeks of 15-minute history.
On a standard server CPU with 8 threads, forecasting one day ahead from three weeks of 15-minute history takes 10 ms with the 5M model and 140 ms with the 209M model.
Source-stated
Mistral says most of Timestral’s training data was synthetic, followed by training on a mix of real and synthetic data.
The majority of the training data was synthetic.
Source and contributor context
Mistral AI · Publisher
The publisher is identified here. No individual author has been inferred.
Source media
Media was declared by the source, but no asset passed the supported availability checks. Open the original source to view it.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Time-series
Observations arranged over time, such as monthly sales; forecasting uses their patterns to estimate later values.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Mistral AI
Mistral AI develops models and products with several options for how organizations deploy them.
Editorial provenance
Version 2. Publication ID: 757bc1fd-1d1e-46c5-891b-bd8e90d29abe.
Prepared by gpt-6-luna; grounding review by server-correction-guard/v1.
Source check: passed by anthropic/claude-opus-5.5 on 2026-10-21T19:01:56.626Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 9e82dc1f1b422b8ae48a53d5b2e7eb8c74297de3a739acd5bdb3046c9f863243. Source text hash: a1835f90ccbe422f35f2db106fe49906ce32a70c1873635c85b261a31aa38140.
Correction reason: Wording revised on 2026-10-21; claims, quotes, numbers and sources are unchanged.
Revision and source-check history
Version 2 · 2026-10-21T12:27:39.637Z · Wording revised on 2026-10-21; claims, quotes, numbers and sources are unchanged.
Mistral AI announces Timestral, a compact family of time-series forecasting models
Mistral AI describes Timestral as its first family of time-series models trained broadly on data to forecast patterns in sequences, such as sensor readings. The four models range from 5 million to 209 million learned numerical settings and produce forecasts across related data series that represent uncertainty. Mistral reports that its largest model ranked first on three public benchmarks, and says the models can run forecasts on a standard server CPU. The company says most training data was synthetic and that a technical report with full results will follow.
Original source
Mistral AI · Article
Source details
Original source date: Oct 15, 2026 (2026-10-15T12:00:00.000Z).
Date basis: Date supplied by the publisher’s feed.
Added to Escape Meridian: 2026-10-21T12:27:39.637Z. First observed: 2026-10-15T15:06:37.276Z. Retrieved: 2026-10-16T02:29:03.772Z.
Limits
Benchmark rankings and performance figures are reported by Mistral in the supplied publisher page; the evidence does not independently verify them. Mistral says results were submitted to the GIFT-Eval and TIME leaderboards and that a technical report with full results will follow.
Open question
How well will Timestral’s reported benchmark performance transfer to real-world forecasting tasks?
Source-stated
Mistral describes Timestral as its first family of time-series foundation models.
Today, we’re releasing Timestral, our first family of time-series foundation models.
Source-stated
Mistral says the four-model family ranges from 5 million to 209 million parameters and supports probabilistic forecasts across multiple related series, with past and known-future covariates as inputs.
Timestral delivers probabilistic, multivariate forecasts and supports past-only and known-future covariates.
Source-stated
Mistral reports that its 209-million-parameter model ranked first on three public forecasting benchmarks; it says the results were submitted to the GIFT-Eval and TIME leaderboards.
Our largest model, at just 209M parameters, ranks first on all three of the field's main public benchmarks, GIFT-Eval, TIME, and fev-bench, outperforming the previous state of the art with only two-thirds of the model parameters.
Source-stated
Mistral says Timestral can run inference on a standard server CPU and gives example runtimes for forecasts one day ahead from three weeks of 15-minute history.
On a standard server CPU with 8 threads, forecasting one day ahead from three weeks of 15-minute history takes 10 ms with the 5M model and 140 ms with the 209M model.
Source-stated
Mistral says most of Timestral’s training data was synthetic, followed by training on a mix of real and synthetic data.
The majority of the training data was synthetic.
Source and contributor context
Mistral AI · Publisher
The publisher is identified here. No individual author has been inferred.
Source media
Media was declared by the source, but no asset passed the supported availability checks. Open the original source to view it.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Time-series
Observations arranged over time, such as monthly sales; forecasting uses their patterns to estimate later values.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Mistral AI
Mistral AI develops models and products with several options for how organizations deploy them.
Editorial provenance
Version 2. Publication ID: 757bc1fd-1d1e-46c5-891b-bd8e90d29abe.
Prepared by gpt-6-luna; grounding review by server-correction-guard/v1.
Source check: passed by anthropic/claude-opus-5.5 on 2026-10-21T19:01:56.626Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 9e82dc1f1b422b8ae48a53d5b2e7eb8c74297de3a739acd5bdb3046c9f863243. Source text hash: a1835f90ccbe422f35f2db106fe49906ce32a70c1873635c85b261a31aa38140.
Correction reason: Wording revised on 2026-10-21; claims, quotes, numbers and sources are unchanged.
Version 1 · 2026-10-16T02:29:40.106Z
Mistral announces Timestral, a compact family of time-series forecasting models
Mistral says Timestral is its first family of time-series foundation models: models trained to forecast patterns in sequences of data, such as sensor readings. The four models range from 5 million to 209 million parameters and produce probabilistic forecasts across related data series. Mistral reports that its largest model ranked first on three public benchmarks, and says it can run forecasts on a standard server CPU. The company says most training data was synthetic and that a technical report with full results will follow.
Original source
Mistral AI · Article
Source details
Original source date: Oct 15, 2026 (2026-10-15T12:00:00.000Z).
Date basis: Date supplied by the publisher’s feed.
Added to Escape Meridian: 2026-10-16T02:29:40.106Z. First observed: 2026-10-15T15:06:37.276Z. Retrieved: 2026-10-16T02:29:03.772Z.
Limits
Benchmark rankings and performance figures are reported by Mistral in the supplied publisher page; the evidence does not independently verify them. Mistral says results were submitted to the GIFT-Eval and TIME leaderboards and that a technical report with full results will follow.
Open question
How well will Timestral’s reported benchmark performance transfer to real-world forecasting tasks?
Source-stated
Mistral describes Timestral as its first family of time-series foundation models.
Today, we’re releasing Timestral, our first family of time-series foundation models.
Source-stated
Mistral says the four-model family ranges from 5 million to 209 million parameters and supports probabilistic forecasts across multiple related series, with past and known-future covariates as inputs.
Timestral delivers probabilistic, multivariate forecasts and supports past-only and known-future covariates.
Source-stated
Mistral reports that its 209-million-parameter model ranked first on three public forecasting benchmarks; it says the results were submitted to the GIFT-Eval and TIME leaderboards.
Our largest model, at just 209M parameters, ranks first on all three of the field's main public benchmarks, GIFT-Eval, TIME, and fev-bench, outperforming the previous state of the art with only two-thirds of the model parameters.
Source-stated
Mistral says Timestral can run inference on a standard server CPU and gives example runtimes for forecasts one day ahead from three weeks of 15-minute history.
On a standard server CPU with 8 threads, forecasting one day ahead from three weeks of 15-minute history takes 10 ms with the 5M model and 140 ms with the 209M model.
Source-stated
Mistral says most of Timestral’s training data was synthetic, followed by training on a mix of real and synthetic data.
The majority of the training data was synthetic.
Source and contributor context
Mistral AI · Publisher
The publisher is identified here. No individual author has been inferred.
Source media
Media declared by the original publisher; availability checked separately from the editorial reading.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Time-series
Observations arranged over time, such as monthly sales; forecasting uses their patterns to estimate later values.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Mistral AI
Mistral AI develops models and products with several options for how organizations deploy them.
Editorial provenance
Version 1. Publication ID: e2a78c3d-0930-49da-80ad-1326c317909a.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-10-16T02:29:40.106Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 9e82dc1f1b422b8ae48a53d5b2e7eb8c74297de3a739acd5bdb3046c9f863243. Source text hash: a1835f90ccbe422f35f2db106fe49906ce32a70c1873635c85b261a31aa38140.
passed · 2026-10-16T02:29:40.106Z · anthropic/claude-opus-5.5
Publication: e2a78c3d-0930-49da-80ad-1326c317909a
attribution — pass: The draft consistently attributes the claims to Mistral, matching the publisher page authored by Mistral.
Today, we’re releasing Timestral, our first family of time-series foundation models.
summary — pass: The title, summary and question accurately reflect the announcement: four model sizes, probabilistic multivariate forecasts, reported benchmark rankings, CPU inference, mostly synthetic training data and a forthcoming technical report. The question about real-world transfer fits the source's own discussion of adaptation.
Timestral has been optimised as a family of four models: 5M, 28M, 123M and 209M parameters.
date — pass: The originalPublished value of 2026-10-15 matches the dateline on the page.
Introducing Timestral October 15, 2026
support — pass: Each claim is directly supported by the source text, including the leaderboard submissions and the two-stage training on synthetic and then mixed data. The draft frames the benchmark results as self-reported.
Timestral is first pre-trained at scale on purely synthetic data with context size 8192, then post-trained on a mix of real and synthetic data with a 4x longer context.
media — pass: The selected media is the verified direct link to the original Mistral announcement, which is the most relevant access point to the source.
passed · 2026-10-21T19:01:56.626Z · anthropic/claude-opus-5.5
Publication: 757bc1fd-1d1e-46c5-891b-bd8e90d29abe
attribution — pass: All claims about the release, model sizes, benchmark rankings, CPU runtimes and training data are attributed to Mistral AI, which publishes the page. Benchmark results are presented as reported by Mistral, not as independently verified.
Today, we’re releasing Timestral, our first family of time-series foundation models.
summary — pass: The title, summary and question accurately reflect the announcement. Terms are explained in plain language: time-series forecasting through sensor readings, parameters as learned numerical settings, and probabilistic forecasts as forecasts that represent uncertainty. The question about whether benchmark results transfer to real tasks follows from the source's own point that production data can differ from benchmarks.
Public benchmarks measure general performance, but production data can be significantly different in behavior and structure.
date — pass: The originalPublished value of 2026-10-15 matches the date shown on the publisher page. The later wording revision does not change the original publication date.
Introducing Timestral October 15, 2026
support — pass: Each claim is supported by the source. The source also supports the statement that pre-training on synthetic data was followed by a real-and-synthetic mix. The limitations correctly note that the results were submitted to leaderboards, are unverified, and will be detailed in a later technical report. No benefits or impacts are added beyond the evidence.
Timestral is first pre-trained at scale on purely synthetic data with context size 8192, then post-trained on a mix of real and synthetic data with a 4x longer context.
media — pass: The selected media is the verified, direct link to the original Mistral announcement page, which is the most relevant way to access the source.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.