Mistral's Timestral runs on CPUs but its Modified MIT license cuts off at $20M monthly revenue

What Timestral is

On October 15, 2026, Mistral released Timestral, a family of four time-series foundation models built for zero-shot forecasting: you hand them a history of numbers and they predict what comes next, with an uncertainty band around the estimate, without any training on your data first. According to a dev.to write-up by Gabriel Anhaia, Mistral claims its largest checkpoint places first across the three major public benchmarks in this domain, and that inference runs without a GPU.

The checkpoints are tiny by current standards. In fp32, the weights span 21 MB for Timestral-5M, 111 MB for the 28M, 490 MB for the 123M and 837 MB for the 209M. The 209M model card, as relayed by dev.to, says it accepts up to 32,768 historical steps and forecasts up to 1,024 steps per forward pass, returning 21 quantiles from 0.01 to 0.99 for every step. It handles multiple series at once, covariates that are either past-only or known into the future (holidays, scheduled promotions), and NaN for missing values.

Speed claims are also Mistral's: on a server CPU with 8 threads, forecasting one day ahead from three weeks of 15-minute history reportedly takes 10 ms with the 5M model and 140 ms with the 209M, with the 28M about five times faster than the largest variant. Timestral 1.0 is listed as a Public Preview with no API pricing announced, so for now the model is download-and-self-host.

The benchmark numbers come from Mistral

Every performance figure originates with Mistral itself, and dev.to notes the accompanying tech report has not been published yet, so the methodology cannot be independently checked. Anhaia also says he has not run the model himself.

On the model card's comparison table, Timestral-209M posts a MASE of 0.658 on GIFT-Eval, ahead of t0-gamma at 0.661, TimesFM-3 at 0.667 and Chronos-2 at 0.698, where lower is better. On fev-bench's SQL skill score it records 50.39 against 49.42 for t0-gamma and 47.28 for Chronos-2. It leads every row, though the margins over the runner-up are narrow.

The smaller models may matter more to most teams. The 28M scores 0.678 on GIFT-Eval MASE and 0.638 on TIME, beating Chronos-2, which Mistral describes as roughly four times larger, but trailing t0-gamma and TimesFM-3 on GIFT-Eval and sitting just behind t0-gamma on TIME. Mistral says the results come from a single checkpoint with no ensembling, submitted to the GIFT-Eval and TIME leaderboards, and that most training data is synthetic, so nothing that was backtested appeared during pre-training. That addresses leakage concerns, but the tech report still has to substantiate it.

The Modified MIT license

The inference code is Apache 2.0. The weights are not. They ship on the Hugging Face Hub under what Mistral calls a Modified MIT license, and the modification is a revenue ceiling: the license withholds authorization from any company whose global consolidated monthly revenue, or that of its employer, exceeded $20 million in the preceding month. The restriction covers derivatives, modifications and combined works, including fine-tuned copies, and even the 5M checkpoint carries the same cap, so a smaller model is not a workaround. Companies above the line can request a commercial license, granted at Mistral's sole discretion, or use its hosted services.

As Anhaia reads it, while noting he is not a lawyer, three consequences follow. The threshold is measured monthly, so a company near the line can be inside the license one month and outside it the next. The restriction follows the weights into anything derived from them. And it counts the whole employer's revenue, however small the using team's budget. There is also a tension he flags: Mistral pitches Timestral for on-prem, VPC and air-gapped deployment, yet the enterprises most likely to want air-gapped forecasting are often precisely the ones above $20 million a month. That is his inference from the two documents, not a Mistral statement.

Practical limits

The package is not on PyPI; it installs from GitHub and requires Python 3.10+ and torch 2.4+. The documented constraints include truncation of context beyond 32,768 steps to the most recent window, block-by-block forecasting for horizons past 1,024 steps with each block's median fed back as context, and covariates that condition forecasts but are never forecast themselves. The project accepts no pull requests and handles issues on a best-effort basis.

Why it matters

Timestral is notable on two fronts. Technically, zero-shot forecasting with probabilistic output at sizes from 21 MB to 837 MB, running on ordinary CPUs, puts predictive capacity planning and anomaly alerting within reach of standard servers rather than GPU clusters. Commercially, the license is the bigger signal: a grant styled after MIT with a revenue ceiling sets a precedent for weights that are free only below a certain business size, and it follows the weights into fine-tunes and combined works. Any team considering adoption should backtest Timestral against its current tooling on real history, and confirm where its employer sits relative to the $20 million monthly line, before building anything on it.