Mistral AI releases Timestral family of time-series foundation models

What happened

Mistral AI has released Timestral, its first family of time-series foundation models designed for zero-shot probabilistic and multivariate forecasting [1]. The model family consists of four different sizes based on parameter counts: 5M, 28M, 123M, and 209M [1].

According to Mistral AI, the 209M parameter model ranks first among pretrained and zero-shot models on three public benchmarks: GIFT-Eval, TIME, and fev-bench [1, 2]. The models support contexts of up to 32K timesteps and can handle both past-only and known-future covariates [2]. They also provide probabilistic quantile outputs to show uncertainty levels [2].

Technical details regarding the model include:

• The architecture uses synthetic pre-training, covariate-aware attention, and long-context learning [1].
• Users can use Mistral Forge for fine-tuning to adapt the models to specific domains [1].
• The models are designed to run on standard server CPUs with 8 threads without needing a GPU [1].
• For a single day of forecasting using three weeks of 15-minute history, the 5M model takes 10 ms, while the 209M model takes 140 ms [1].
• The 28M model is five times faster than the 209M model, and the 123M model reaches within 0.2% of the 209M score while using three quarters of its latency [1].
• The 28M model outperformed Chronos-2 on GIFT-Eval and TIME, while the 209M model outperformed TimesFM-3 and Falcon-2.0 despite having fewer parameters [1].

Timestral can process categorical covariates, such as the day of the week or machine states like ‘idle’ or ‘running’, directly as inputs [1]. It is a numerical model that can be integrated into monitoring pipelines or used alongside Mistral language models in multi-modal agentic workflows [1].

Who is affected

This technology is intended for various professional sectors including:

• Industrial plants: For forecasting process sensors against production schedules, detecting equipment drift, and planning spare parts [1].
• Finance teams: For forecasting volumes, flows, and volatility to assist with treasury, hedging, and positions across thousands of accounts [1].
• Enterprise users: For managing sensor telemetry and observability metrics [1].