Mistral Timestral forecasts on a CPU. Your license is probably fine.
Mistral's Timestral time-series models top forecasting benchmarks, run on a CPU, and cap free use at $20M monthly revenue. Most small businesses clear it.
In September we told you Google's TimesFM-3 topped every forecasting leaderboard and then shipped under a non-commercial license. Mistral Timestral is the follow-up worth reading. Mistral's first time-series models claim the same leaderboards, run on a plain server CPU, and the license only stops you if your company makes more than $20 million a month.
What actually happened
Mistral released Timestral on October 15: four forecasting models at 5M, 28M, 123M and 209M parameters, same architecture and training recipe. Mistral says the 209M model ranks first among pre-trained zero-shot models on GIFT-Eval, TIME and fev-bench, ahead of TimesFM-3 with fewer parameters, and that the 28M model beats Chronos-2, a model about four times its size.
The practical specs are the real story:
- No GPU needed. On an 8-thread server CPU, a one-day forecast from three weeks of 15-minute data takes 10 ms on the 5M model and 140 ms on the 209M.
- Covariates built in. Past-only and known-future inputs, including categorical ones — promotions, holidays, price changes.
- Probabilistic output. 21 quantiles, so you get a range, not a single guess.
- 32,768-step context.
The code is Apache 2.0 on GitHub. The weights on Hugging Face use a modified MIT license: rights don't apply if your company's (or employer's) global monthly revenue exceeded $20 million in the prior month, and that cap follows derivatives and fine-tunes (license). Above it, you ask Mistral for a commercial license.
Why Mistral Timestral matters for your business
For a small operator, this flips the TimesFM-3 answer from "no" to "yes." $20M a month is $240M a year. If you're reading this blog, you're very likely under it.
What we'd do with it:
- Start with 28M, not 209M. It's about five times faster and Mistral's own numbers put it ahead of much bigger models. Forecasting SKUs weekly doesn't need the top checkpoint.
- Use the quantiles. Order to the 70th percentile on fast movers and the 50th on slow ones. A range turns a forecast into a purchasing rule.
- Keep the model behind an interface. Same rule as last time. The pipeline — clean sales history, renamed SKUs, promotion flags — is the asset. If you grow past the cap, you swap the model, not the system.
- Write the license check into your vendor list. Revenue caps follow the company, including an acquirer.
Key takeaways
- Mistral Timestral ships four time-series models: 5M, 28M, 123M and 209M parameters
- Mistral reports first place among zero-shot models on GIFT-Eval, TIME and fev-bench
- Runs on a CPU: 10–140 ms per one-day forecast on an 8-thread server
- Weights are modified MIT: no rights for companies above $20M global monthly revenue, including for derivatives
- Start with the 28M model, use the quantiles, and keep the model swappable
A forecast should end in a purchase order. We build demand forecasting on your own servers, on licenses you're allowed to use, wired into how you actually buy. Estimate what better ordering is worth, or see how we build it.
Sources: Mistral AI, Hugging Face license, GitHub.
- #forecasting
- #time-series
- #mistral
- #model-licensing
- #inventory
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
Microsoft calls AI agents insiders: start them at zero access
Microsoft's agentic insider push adds Dynamic Permissions and Dynamic Agent Risk. The lesson works without an E7 license: give AI agents zero standing access.
Read itMeta: Claude tool tripped No Kings Instagram lockout. Use the API
Meta says a Claude-based tool reading Instagram at 4x normal speed triggered the No Kings lockout. AI agents on your business accounts can trip bot detection.
Read it