Mistral published Timestral, time-series models from 5M to 209M

On 15 October 2026, Mistral published Introducing Timestral, its first family of time-series foundation models. The post is dated that day. The Timestral 1.0 docs label the release Public Preview.

The family has four sizes: 5M, 28M, 123M, and 209M parameters. Mistral says all four share the same architecture and training recipe. The spec table on the post lists a context of 32,768 time steps, probabilistic forecasts with 21 quantiles, and inputs of multivariate targets plus past-only and known-future covariates. The GitHub README states the same three points: past-only, known-future, and categorical covariates; 21 quantiles per time step; context up to 32,768 steps, with the most recent steps kept when a series is longer.

The docs navigation lists Timestral 1.0, whose weights row shows 209M parameters, plus separate pages for 123M, 28M, and 5M. Each of those pages carries the Public Preview and Modified MIT labels. The 209M model card says a single forward pass forecasts up to 1,024 steps, and that longer horizons are forecast block by block. It lists 21 quantiles from 0.01 through 0.99, and it describes a patch-based transformer with 24 layers and an embedding dimension of 768.

What Mistral says on the benchmarks

Mistral says the 209M model ranks first among pretrained, zero-shot models on GIFT-Eval, TIME, and fev-bench, across five metrics, from a single checkpoint, without ensembling. The post says the results have been submitted to the GIFT-Eval and TIME leaderboards. The fev-bench chart is captioned as leaderboard entries. Those ranks and the submission line are Mistral’s. The leaderboard sites were not re-opened for this account, and the scores were not rerun here.

One label differs between Mistral’s own pages. The post’s GIFT-Eval charts name geometric-mean MASE and geometric-mean WQL. The 209M card’s comparison table lists MASE and CRPS for GIFT-Eval and for TIME, plus SQL and MASE skill scores for fev-bench. Both presentations are Mistral’s. The card’s 209M figures include GIFT-Eval MASE 0.658 and CRPS 0.446, TIME MASE 0.626 and CRPS 0.522, and fev-bench SQL skill 50.39. Treat that table as the vendor’s.

The post also says Timestral does not need a GPU at inference. On a standard server CPU with 8 threads, forecasting one day ahead from three weeks of 15-minute history takes 10 ms with the 5M model and 140 ms with the 209M model. Mistral adds that the 28M model is five times faster than the 209M model, and that the 123M model comes within 0.2% of the 209M score at three quarters of its latency. Those timings are Mistral’s measurements, not an independent clock.

On training, Mistral says most of the data was synthetic, that pre-training used a context of 8,192, and that a later stage mixes real and synthetic data at a context four times longer. The objective it names is a pinball loss over 21 quantiles. Four times 8,192 is 32,768, the context length in the spec table. A technical report is not up yet: the post says it will follow, and the model card says it is coming soon.

Two licenses

The repository LICENSE is the Apache License, Version 2.0, and the README says that grant covers the code. The weights are separate. The 209M LICENSE is a Modified MIT license. The same text is on the 5M, 28M, and 123M weight repos. It keeps the attribution notice, then adds a revenue condition: you are not authorized to exercise the rights if the global consolidated monthly revenue of your company, or of your employer, exceeds $20 million (or the equivalent in another currency) for the preceding month. The condition applies to the model and to derivatives, modifications, and combined works. Above that line, the text says you may ask Mistral for a commercial license, which Mistral may grant at its sole discretion, or use the model on Mistral’s hosted services. The revenue condition is Mistral’s addition. The Apache grant on the code does not carry it.

The Hub API records for 209M, 123M, 28M, and 5M, read on 16 October 2026, list each repo as public (gated false), license: other, and last modified on 15 October 2026.

Adapting it

Mistral says Timestral is available through Mistral Forge, the training pipeline it uses for its language models. The post lists zero-shot inference, architectural extensions such as domain-specific covariates, fine-tuning on proprietary data, and control of the training mix, including replay of pre-training data and synthetic samples. It says the full pipeline can be deployed on a client’s hardware, including air-gapped environments.

For a forecast you can run on CPU, the weights and the Apache-licensed loader are public. Read the revenue condition before you ship a product on them, and treat the leaderboard rank and the 10 ms and 140 ms figures as Mistral’s until you measure your own series.