Andy Garcia · freelance dev
FRENES
Back to the blog
Open source

Mistral's Timestral Forecasting Models Run on a CPU, No GPU Needed. The 'Modified MIT' Licence Stops at $20 Million a Month in Revenue.

October 19, 2026 7 min read

On Thursday, October 15, Mistral released Timestral, its first family of time-series foundation models. Four sizes: 5M, 28M, 123M and 209M parameters. The weights are on Hugging Face, the inference code is on GitHub, and Mistral says the largest model ranks first among pretrained, zero-shot models on the field's three main public benchmarks, GIFT-Eval, TIME and fev-bench. It needs no GPU. On a standard server CPU with 8 threads, Mistral measures 10 ms for the 5M model and 140 ms for the 209M model to forecast one day ahead from three weeks of 15-minute history.

The launch post gives the licence one line in a spec table: "Modified MIT". The word that matters is "Modified".

I am not a forecasting specialist. I am the CTO of a LegalTech built on Claude, and I built BidForge, an AI that reads French public tender files for small businesses. In both jobs I read the licence before the benchmark, because the licence is the one part of a model you cannot fine-tune. What follows is my reading of a short text, not legal advice.

What Mistral shipped

A time-series foundation model does for sensor readings, sales and cash flows what a language model does for text: it is trained once on a large corpus, then asked to forecast series it has never seen, without training on your data. Timestral produces probabilistic, multivariate forecasts: 21 quantiles per time step, from 0.01 to 0.99, so you get a range and not just a line. It takes past-only and known-future covariates, such as a production schedule or a holiday calendar, and categorical ones, such as a machine's state.

The context goes up to 32,768 time steps. At 15-minute resolution, that is 32,768 / 96, about 341 days of history. One forward pass forecasts up to 1,024 steps. On disk, the four checkpoints weigh 21 MB, 111 MB, 490 MB and 837 MB.

The speed matters more than it sounds. Three weeks of 15-minute data is 2,016 points, and one day ahead is 96. Say you forecast 5,000 product series every night, one call per series, one after the other, and Mistral's CPU figures hold on your machine. At 140 ms, that is 700 seconds, under 12 minutes, on 8 threads. Mistral says the 28M model is five times faster than the 209M, so about 140 seconds. The 5M does it in 50 seconds. No GPU budget, no API bill.

The clause the GitHub badge does not show

GitHub displays Apache 2.0 for the repository, the package metadata declares Apache-2.0, and that is the licence a scanner reading the package is likely to report. But the README says the code and the weights are licensed separately. The weights, which from_pretrained downloads into the Hugging Face cache on first use, come under the "Modified MIT" licence.

It starts like MIT: free of charge, use, copy, modify, sell. Then condition 2:

"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."

The restriction, the licence adds, "applies to the Model and any derivatives, modifications, or combined works based on it, whether provided by Mistral AI or by a third party." Above the line, you can request a commercial licence at sales@mistral.ai, "which Mistral AI may grant you at its sole discretion", or use Mistral's hosted services.

All four Timestral sizes carry it. The same text, word for word apart from the year in the attribution line, sits on Devstral 2, the 123B coding model Mistral released on December 9, 2025 and described as shipping "under a modified MIT license". In both files, the clause even calls itself "(b)" while being numbered 2.

Who is in, who is out

$20 million a month is $240 million a year. For the small businesses I build for, this licence behaves like MIT. Four details decide which side of the line you are on.

It is a monthly test. The licence looks at the preceding month, not the year. A company that averages $15 million a month but books $25 million in December is over the line in January, as I read it.

It counts your employer. An engineer at a company above the line cannot use the weights under this licence, not even to evaluate them. The text has no exception for research or testing.

It follows the model. A fine-tuned Timestral that a consultant installs at a large client is still covered. If you sell an on-prem forecasting tool, your client's revenue matters too, not only yours. If you host the model yourself and sell forecasts, as I read it, the question is your own revenue, until the month it is not.

It says "consolidated" without defining it. The text does not say whether a subsidiary counts its parent's revenue. If you build a startup on these weights, ask Mistral before an acquirer's lawyers ask you.

The hosted path has no published price either. On October 19, Mistral's pricing page had no Timestral line. The model's docs page lists weights and a licence, labels the model "Public Preview" and shows no API price. The Forge page, through which Mistral offers fine-tuning, ends on "Talk to an expert".

Mistral's own showcase is an industrial plant with thousands of sensors, and its spec sheet lists on-prem, VPC or air-gapped deployment. I would expect many companies that run plants like that to sit above $240 million a year. For them, Timestral is a sales conversation, not a download.

What the benchmark lead is worth

The numbers are Mistral's own, and the technical report is still "coming soon" on the model card. There is one outside check. Mistral submitted its GIFT-Eval results with a notebook that lets anyone rerun them, and on October 16 the Salesforce AI Research scientist who merged them wrote: "I have reproduced identical results across all model sizes." TIME's maintainers merged Mistral's results the same day with a "LGTM" and no mention of a rerun. That is better than a chart in a blog post. A rerun confirms the numbers, not that they will hold on your data.

The model card's table is the useful part, because it lists the runners-up. On GIFT-Eval, Timestral-209M scores 0.658 MASE, where lower is better. Second is t0-gamma, from The Forecasting Company, at 0.661: a gap of 0.45 percent. Across the four error scores of GIFT-Eval and TIME, the widest gap to t0-gamma is TIME's CRPS, 0.522 against 0.534, about 2.2 percent. On fev-bench, which reports skill scores, the lead is 1 to 1.5 points. Chronos-2, Amazon's 120M model, scores 0.698 on GIFT-Eval MASE. The GIFT-Eval leaderboard, read on October 19, shows the same GIFT-Eval figures and ranks Timestral-209M first among pretrained and zero-shot entries. It also lists one pretrained entry with a lower MASE, 0.646: JPMorgan Chase's Morpheus-2.0, added the same night, with no public weights and a worse CRPS, 0.455.

Now the licences, all read on Hugging Face on October 19. t0-gamma, about 256M parameters: Apache 2.0. Chronos-2: Apache 2.0. Datadog's Toto 2.0: Apache 2.0. Google's TimesFM 3.0: "TimesFM Non-Commercial License v1.0".

So a company above the line chooses between a sales call with no public price and an Apache 2.0 model that, on Mistral's own table, trails by about 2.2 percent at most on those error scores. Everyone below the line gets a free, fast, strong model, and per Mistral, the 28M version, at 111 MB, beats Chronos-2 on GIFT-Eval and TIME.

What I would check before building on it

Write down whose revenue counts. Yours, your employer's, and, for every on-prem deployment, your client's. Make it a line in the contract, not an assumption.

Put the weights in your licence inventory. Pin the Hugging Face revision (from_pretrained takes a revision), keep the LICENSE file next to the weights, and tag them "Modified MIT" by hand.

Backtest on your own data against an Apache 2.0 model. Run Timestral and Chronos-2 or t0-gamma on the same history and compare the quantiles, not just the median. Mistral itself writes that production data can differ from benchmarks and shift over time. If the gap on your data is as small as on the public tables, the licence is the tiebreaker.

Keep the forecaster behind one interface. If your company or your client crosses $20 million in a month, swapping models should be a configuration change, not a rewrite.

And wait for the technical report before you say "state of the art" to a client. Mistral says most of its training data is synthetic. The report is where you will see what that means.

For a small business, Timestral is a free forecaster that runs on the server it already has. For a large one, it is an email to sales. The difference is one paragraph of the licence, and it is the second one.

Sources

A project like this one?

I design and deploy products like this. Let's talk.

Let's talk