RC RANDOM CHAOS

Timestral Beats All Time-Series Benchmarks at Half the Parameters

· via Hacker News

Original source

Timestral: Mistral's time-series foundation models

Hacker News →

Mistral released Timestral, a family of four time-series foundation models ranging from 5M to 209M parameters. The largest variant ranks first across all three major forecasting benchmarks—GIFT-Eval, TIME, and fev-bench—while using only two-thirds the parameters of the previous state-of-the-art. This represents a departure from classical approaches like ARIMA or deep learning models trained on individual datasets, instead leveraging pre-training on diverse data to enable zero-shot forecasting across new domains and data-scarce scenarios.

The models deliver multivariate probabilistic forecasts that handle both continuous and categorical covariates, including known-future information like production schedules or weather forecasts. They run efficiently on standard CPUs without GPU requirements, with inference latencies from 10ms for the smallest model to 140ms for the 209M variant. The architecture employs decoupled horizon forecasting, covariate-aware attention mechanisms, and a synthetic-first training strategy, followed by a two-stage process that pre-trains at scale on synthetic data before fine-tuning on a mixed real-and-synthetic dataset.

In enterprise settings, Timestral targets industrial monitoring, supply-chain forecasting, and financial planning. The system can monitor sensor streams for anomalies, then integrate with language model agents to perform root-cause analysis by matching patterns against historical data and drawing on contextual information like maintenance logs and equipment manuals. Domain-specific adaptation is supported through Mistral Forge fine-tuning.

Read the full article

Continue reading at Hacker News →

This is an AI-generated summary. Read the original for the full story.