On October 15 Mistral released Timestral, its first family of time-series foundation models, in 5M, 28M, 123M and 209M sizes. Weights are on Hugging Face under a Modified MIT license and the inference code is Apache 2.0. The models produce multivariate probabilistic forecasts with 21 quantiles per step, accept past-only and known-future covariates, and take up to 32,768 steps of context. Mistral says the 209M model ranks first among pretrained zero-shot models on GIFT-Eval, TIME and fev-bench with about two-thirds the parameters of the previous best; on an 8-thread CPU, forecasting one day from three weeks of 15-minute data takes 10 ms with the 5M model and 140 ms with the 209M. These are Mistral's own figures; the technical report is still to come.
Key Takeaways
- ✓Four sizes, 5M / 28M / 123M / 209M, with fp32 weights from 21 MB to 837 MB, all downloadable from Hugging Face.
- ✓Mistral says the 209M ranks first among pretrained zero-shot models on GIFT-Eval, TIME and fev-bench with about two-thirds the parameters of the prior best (vendor-reported).
- ✓The 28M beats Chronos-2, about 4x larger, on GIFT-Eval and TIME; the 209M beats TimesFM-3 and Falcon-2.0 with fewer parameters (vendor-reported).
- ✓No GPU needed: a one-day forecast on an 8-thread CPU takes about 10 ms (5M) or 140 ms (209M); the 123M lands within 0.2% of the 209M score at three-quarters of its latency.
- ✓32,768-step context, up to 1,024 steps per forward pass, 21 output quantiles; fine-tuning and on-prem deployment via Mistral Forge.
Key Decision Metrics at a Glance
Turn your technical choice into a development budget
Compare 40 dev plans & simulate token costs vs $20/mo subscriptions
Project Links & Resources
Direct AccessIn-Depth Technical Analysis
Background
Sensor telemetry, observability metrics, sales and cash flows are all time series that drive capacity planning, energy and maintenance decisions. Classical methods such as ARIMA fit one series at a time, and deep models like PatchTST usually train from scratch per dataset. Time-series foundation models pre-train on large, diverse data and forecast zero-shot on unseen series. Timestral is Mistral's first model outside language.
How it works
Timestral is a causal, patch-based transformer. Mistral highlights Decoupled Horizon Forecasting, where the full history stays in the sequence and separate forecast tokens read it; covariate-aware attention, where known-future covariates such as holidays or weather forecasts are visible across the window while targets stay causal; mostly synthetic training data; and two-stage training, first on synthetic data at 8192 context, then on real plus synthetic data at 4x the context, using a pinball loss over 21 quantiles.
Benchmarks
Per Mistral, the 209M ranks first among pretrained zero-shot models on GIFT-Eval, TIME and fev-bench with about two-thirds the parameters of the previous best; the 28M beats Chronos-2 (about 4x larger) on GIFT-Eval and TIME, and the 209M beats TimesFM-3 and Falcon-2.0. Results were submitted to the GIFT-Eval and TIME leaderboards; the technical report and independent replication are still pending.
Getting started
Install the timestral package from GitHub with pip and load weights with TimestralForecaster.from_pretrained. Inputs can be NumPy arrays or torch tensors with NaN for missing values, and outputs include the median and any quantile. It runs on CPU, so it drops into existing monitoring pipelines; pick 5M or 28M for latency and 123M or 209M for accuracy. Mistral Forge covers fine-tuning and air-gapped deployment.
Benchmark side-by-side against alternatives, or calculate monthly token cost vs subscription break-even.
Discussion & Comments
0Sign in to join the discussion
Connect with AI developers to exchange benchmark insights.