Blog
Mistral releases Timestral, small open forecasting models that run on a CPU
On 15 October 2026 Mistral released Timestral, its first family of time-series foundation models, in four sizes from 5M to 209M parameters under a modified MIT licence. The models forecast sensor, sales and other data zero-shot and run on ordinary CPUs. Benchmark rankings are Mistral's own, and some margins are small.
Mistral released Timestral on 15 October 2026, its first family of time-series foundation models. These are AI models that forecast numbers over time, such as sensor readings, sales volumes or energy demand, rather than generating text. The weights are free to download from Hugging Face, and Mistral says the models run on an ordinary CPU, with no GPU needed.
What Timestral is
Time-series data is one of the most common kinds of business data. Factories track sensors, finance teams track cash flows and volumes, and IT teams track server metrics. Classic forecasting tools such as ARIMA are usually fitted to one series at a time and learn only from that series' history.
A time-series foundation model takes the approach used for language models. It is pre-trained once on a large and varied set of data, and can then forecast data it has never seen, which is called zero-shot forecasting. Mistral says this helps most in areas where there is little history to learn from.
The models in numbers
According to Mistral's announcement and model card:
- Four sizes: 5 million, 28 million, 123 million and 209 million parameters. That is tiny next to modern language models, which run to billions or trillions of parameters.
- Long memory: each model can read up to 32,768 past time steps and forecast up to 1,024 steps in one pass. Longer forecasts are built block by block.
- Ranges, not single numbers: the models output 21 quantiles, a spread of possible values, so teams can see how uncertain a forecast is.
- Extra inputs: forecasts can take in related series and covariates, meaning outside factors. These can be known only for the past, or known in advance, such as a production schedule, a holiday calendar or a weather forecast. Categories such as "running" or "under maintenance" are also accepted.
- Speed: on a standard server CPU with 8 threads, Mistral says forecasting one day ahead from three weeks of 15-minute readings takes 10 milliseconds with the smallest model and 140 milliseconds with the largest.
- Licence: the weights are released under a modified MIT licence.
Mistral's benchmark claims
Mistral says Timestral-209M ranks first among pre-trained, zero-shot models on the three main public forecasting benchmarks: GIFT-Eval, TIME and fev-bench. It says the 209M model beats Google's TimesFM-3 with fewer parameters, and that the 28M model beats Amazon's Chronos-2, which is about four times larger, on two of the three tests.
These are Mistral's own figures. The company says it has submitted results to the GIFT-Eval and TIME leaderboards. Some margins are small: on the model card, Timestral-209M's GIFT-Eval error score is 0.658, against 0.661 for the next model listed. MarkTechPost also notes that some rival scores in the comparison come from Mistral's model card rather than from Google's or Amazon's own pages.
How it was built
Mistral says most of the training data was synthetic, meaning computer-generated series with trends, seasonality and noise, combined so that related series affect each other. The models were first trained on synthetic data only, then on a mix of real and synthetic data with a longer context. The company also describes a new way of handling forecast targets, which it calls Decoupled Horizon Forecasting, and an attention design that stops a model from seeing a target's future values by accident. Mistral says a full technical report will follow.
Why it matters
Timestral moves Mistral beyond chatbots and language models into numerical forecasting, a core tool for industry and finance. The small size and CPU-only inference mean companies could run forecasts inside existing systems without buying GPUs.
For businesses that need more, Mistral offers fine-tuning on private data through its Mistral Forge platform, including on-premises and air-gapped setups. Mistral's documentation lists the 209M model as a public preview. As with any forecasting tool, results on a company's own data may differ from public benchmark scores.