Mistral releases Timestral forecasting models that run on CPUs

Mistral released Timestral, a family of four models for forecasting numerical data over time, and published a GitHub repository for running them on users' own data. The models range from 5M to 209M parameters and can forecast related series together while estimating uncertainty and incorporating known future inputs, such as production schedules. Mistral says an 8-thread server CPU can forecast one day ahead from three weeks of 15-minute history in 10 ms with the 5M model and 140 ms with the 209M model. Through Mistral Forge, customers can fine-tune the models and deploy the full training pipeline on their own hardware, including air-gapped systems. The company cautions that production data can differ substantially from public benchmarks and shift over time.