Mistral Timestral: four open-weight time-series models, a $20M revenue cap, and a CERN statement of intent

Mistral Timestral is the French lab's first family of time-series foundation models, announced on October 15, 2026. The four models, at 5 million, 28 million, 123 million and 209 million parameters, forecast numbers over time, such as sensor readings, demand or trading volumes, rather than generating text. The weights are downloadable, which puts Timestral among the open models we track, but under a license that excludes large companies. The same day, CERN announced a separate statement of intent with Mistral to build an AI model for particle physics. Both stories reached the top of Hacker News on October 19.

This story is based on Mistral's announcement, model cards, license file, GitHub README and docs, CERN's announcement and Reuters' report, checked October 20, 2026. For how Mistral's license compares with Apache, MIT and Llama terms, see our open-weight licenses guide.

The short version

• What: four Timestral models (5M, 28M, 123M, 209M) sharing one architecture.
• What they do: zero-shot probabilistic forecasting of one or many related series, with optional covariates.
• License: weights under a "Modified MIT" license with a $20 million monthly revenue cap; code under Apache 2.0.
• Where: Hugging Face (mistralai/Timestral-5M to -209M), a GitHub package, and Mistral Forge for fine-tuning. Mistral hasn't announced a hosted API price.
• Benchmarks: vendor-reported. Mistral says the 209M model ranks first among pretrained zero-shot models on GIFT-Eval, TIME and fev-bench. Its technical report is "coming soon."
• CERN: a statement of intent for a "long-term strategic partnership," not a contract. No money or timeline was disclosed.

What is a time-series foundation model for?

A time-series model takes a history of measurements, such as 15-minute readings from a pump, and predicts what comes next. A "foundation" model is pretrained on many kinds of series, so it can forecast new data without training on it first ("zero-shot").

Mistral pitches Timestral at industrial and financial work: forecasting thousands of sensors (vibration, temperature, pressure) to spot anomalies and plan maintenance, and forecasting flows, volumes and volatility for treasury and hedging.

Timestral outputs probabilities, not a single line: 21 quantiles per step, so you get a range of likely values. It accepts several target series at once, "past-only" covariates (inputs you only know historically) and "known-future" covariates (for example, a holiday calendar).

Timestral at a glance

Last verified: October 20, 2026

What is the license, and is it open source?

The weights can be downloaded, modified and used commercially, with one large exception. The license file on each Hugging Face model says 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" in the preceding month. The cap applies to derivatives and combined works too. Above it, you can request a commercial license, which Mistral "may grant you at its sole discretion," or use its hosted services.

That is the same revenue-gated "Modified MIT" pattern Mistral used for Mistral Medium 3.5, covered in our license comparison. It makes Timestral "open weights," not open source in the OSI sense. The two models Mistral compares against take different routes: Amazon's Chronos-2 (about 120 million parameters) is Apache 2.0 on Hugging Face, while Google's TimesFM 3.0 (0.3 billion parameters) uses the TimesFM Non-Commercial License v1.0.

What do the benchmarks show?

All results below are vendor-reported by Mistral. They haven't been described in a published paper yet.

Last verified: October 20, 2026 (Hugging Face model card)

MASE scores point forecasts; CRPS scores the whole predicted distribution. GIFT-Eval, built by Salesforce AI Research in 2024, spans 23 datasets, about 144,000 series, seven domains and ten frequencies.

Mistral's claims on top of these numbers:

• the 209M model "ranks first" among pretrained zero-shot models on all three benchmarks, with a single checkpoint and no ensembling;
• the 28M model "beats Chronos-2, about four times larger, on GIFT-Eval and TIME";
• the 209M model "beats TimesFM-3 and Falcon-2.0 with fewer parameters."

Limits: Mistral's blog shows rivals' scores only in charts, and its claim that "nothing we backtest was seen during pre-training" can't be checked until the technical report is out. Rankings also move fast: Amazon said Chronos-2 ranked first among pretrained models on GIFT-Eval when it launched in October 2025.

What did CERN and Mistral actually sign?

A statement of intent, CERN's announcement says, "to establish a long-term strategic partnership." It's a declaration of a plan, not a contract with money attached. CERN's page names no signatories, budget, compute commitment or end date, and Reuters reported that no budget or timeline was disclosed.

The plan is to "develop a single pretrained Large Physics Model," which CERN calls the first foundation model for collider physics. It would be trained on billions of simulated proton-proton collisions at the Large Hadron Collider and handle tasks now spread across chains of specialized algorithms: reconstructing collisions, identifying particles, flagging unexpected signals and evaluating detector designs.

The scale behind it: CERN holds about one exabyte of LHC data, and the High-Luminosity LHC, expected around 2030, would raise collision numbers tenfold. CERN's CIO, Enrica Porcari, said the work should "serve the whole physics community and be shared." Reuters called the model "Flare"; CERN's page doesn't use that name.

Neither side has said whether the physics model's weights will be released. Timestral isn't part of the CERN project.

What's confirmed, reported and uncertain

• Confirmed (Mistral): four sizes, 32,768-step context, 21-quantile output, covariate support, the Modified MIT license with its $20M cap, Apache 2.0 code, Hugging Face and GitHub availability, "Public Preview" status.
• Vendor-reported: benchmark scores and rankings, CPU latency, and comparisons with Chronos-2, TimesFM-3 and Falcon-2.0.
• Confirmed (CERN): a signed statement of intent, the Large Physics Model goal, and the LHC data figures.
• Reported (Reuters): the "Flare" name; no budget or timeline disclosed.
• Uncertain: hosted API pricing for Timestral, the technical report, and whether the CERN model will have open weights.

What it means for you

If you forecast operational or financial data and your company is under the $20 million monthly revenue line, Timestral is a small, CPU-friendly option. The 5M model's file is just 21 MB, far smaller than even the laptop language models we track. Above the line, Chronos-2 is the only model Mistral compares against that is under Apache 2.0. Mistral's October language-model release, Mistral Large 4, is covered in our DeepSeek V4.1 Flash story.