Mistral Releases Timestral: Small Time-Series Models for Forecasting on a CPU

Mistral [released Timestral on October 15](https://mistral.ai/news/timestral/), its first family of **time-series foundation models**. The four downloadable checkpoints have **5M, 28M, 123M and 209M parameters**. They take numerical history—not a text prompt—and produce probabilistic forecasts. It is a new model family for forecasting, **not** a replacement for a coding assistant or a general-purpose LLM.

What does it predict?

Timestral can forecast multiple related signals together and incorporate both historical inputs and inputs known in advance, such as a maintenance schedule or calendar. It returns **21 quantiles** for each future time step, making a range of possible outcomes available instead of only a single point estimate. Mistral gives industrial sensors, capacity planning, energy and supply chains as possible domains. These are use cases, not verified deployment outcomes. [Source: Mistral release](https://mistral.ai/news/timestral/)

For a developer, a useful first experiment could be forecasting the next day of request volume from historical traffic and a known holiday calendar, then checking whether the predicted intervals cover actual values. Compare that result against a seasonal baseline and a model trained only on your own history. Time-series evaluation needs a chronological holdout: random train/test shuffling can leak future patterns.

How can you try it, and what does it cost?

Mistral [published inference code](https://github.com/mistralai/timestral) and [four model checkpoints](https://huggingface.co/collections/mistralai/timestral). Its README shows a Python TimestralForecaster.from_pretrained(...) example and says inference works on CPU or GPU. The company reports that a standard eight-thread server CPU can produce a specific one-day forecast in **10 ms with the 5M model** or **140 ms with the 209M model**; those are vendor measurements for that workload, not a latency promise for yours. [Sources: Mistral release](https://mistral.ai/news/timestral/); [GitHub README](https://github.com/mistralai/timestral)

There is **no published Timestral per-token API price** in the release. Downloadable weights avoid a vendor per-token fee for local inference, but hardware, storage and operations are still costs. The repository's code is **Apache 2.0**; the model weights are separately licensed under **Modified MIT**, so read the relevant model license before production use. Mistral also describes adaptation through Forge, but that is a separate organizational route rather than a requirement for running the published checkpoints. [Sources: GitHub README](https://github.com/mistralai/timestral); [Mistral release](https://mistral.ai/news/timestral/)

How strong are the results?

Mistral says the 209M model ranks first among pretrained, zero-shot models on **GIFT-Eval, TIME and fev-bench** in its evaluation and notes that it submitted results to two public leaderboards. The comparison is constrained to the stated benchmark entries and setup; sAIgnals has **not independently reproduced it**. More importantly, benchmark rank cannot tell you whether the model handles your own missing data, unusual seasonality or changing operating conditions. [Source: Mistral release](https://mistral.ai/news/timestral/)

**Bottom line:** Timestral gives developers an unusually compact, locally runnable forecasting option with uncertainty estimates. Start with the 5M or 28M checkpoint for a low-cost prototype, but keep a simple baseline and test on later, untouched data before trusting a production alert or planning decision.