Timestral is a new family of time-series foundation models. Its largest model has 209M parameters and ranks first on GIFT-Eval, TIME, and fev-bench. The article says this beats the previous state of the art while using only two-thirds of its parameters. Time-series forecasting has mostly relied on models trained for one dataset at a time, so a pre-trained model that forecasts well on data it has never seen is a significant change.

Why pre-trained forecasting matters

Time series is one of the most common data types in enterprise, from sensor telemetry to observability metrics. Forecasts built on it drive decisions in capacity planning, energy, supply chains, risk, and maintenance. Industrial plants forecast process sensors against production schedules, catch equipment drift before it fails, and plan capacity and spare parts across thousands of assets. Finance teams forecast flows, volumes, and volatility to inform positions, hedging, and treasury.

Classical models such as ARIMA and linear regression are well established, but they are usually fitted to a single series and learn only from that series' history. Deep learning models such as PatchTST can learn nonlinear patterns across many related series, but they are typically trained from scratch on each dataset, so what they learn does not carry over to new domains. Time series foundation models (TSFMs) apply the pre-training approach to time series. Trained on a large and diverse corpus, they learn patterns that transfer across series and can forecast zero-shot on data they have never seen, including in domains where data is scarce.

Benchmark results

The team evaluated Timestral on GIFT-Eval, TIME, and fev-bench. Among pre-trained, zero-shot models, it ranks first across all five metrics. These results come from a single checkpoint, without ensembling. The results have been submitted to the GIFT-Eval and TIME leaderboards. The article also notes that its model family sits on the Pareto front of forecasting accuracy versus model size, meaning that for each model size, no public model forecasts better.

A family of four sizes

Timestral comes in four sizes: 5M, 28M, 123M, and 209M parameters. Larger models improve forecast performance, while smaller ones are faster and cheaper. All four share the same architecture and training recipe. The 28M model beats Chronos-2, which is about four times larger, on GIFT-Eval and TIME. The 209M model beats TimesFM-3 and Falcon-2.0 with fewer parameters.

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. The 28M model is five times faster than the 209M model. The 123M model comes within 0.2% of the 209M score at three quarters of its latency.

An industrial example

The article illustrates the model with an industrial plant that has thousands of sensors. Timestral improves forecasts in four ways:

• Multivariate forecasting. A pump's vibration, temperature, and pressure are related. Timestral models related series jointly, so it can use those dependencies instead of forecasting each sensor in isolation.
• Probabilistic forecasting. Timestral predicts a full set of quantiles, so teams can see how wide the uncertainty is, set alert thresholds that reflect it, and plan maintenance around risk rather than a single number.
• Categorical covariates. Inputs such as the day of the week or a machine's state (running, idle, under maintenance) are taken as categorical inputs directly, rather than encoded as numbers that would imply an order or scale they lack.
• Known-future covariates. Production schedules, planned shutdowns, and weather forecasts can be used as conditions, so a scheduled load increase is expected rather than flagged as a surprise.

Timestral is a numerical model that can run inside existing monitoring pipelines or be paired with Mistral's language models in multi-modal agentic workflows. In the plant example, Timestral runs continuously on sensor streams and detects anomalies. When it flags a pump, an agent compares the signals against embeddings of past sensor data to find similar historical failures. Sub-agents then run a root cause analysis using maintenance records, shift logs, operator notes, equipment manuals, past work orders, and readings from related equipment. The agent drafts a suggested diagnosis with supporting evidence and possible next steps, so the plant manager receives a diagnosis and plan backed by sensor data rather than just an alert.

Key technical contributions

The article credits four ideas for the results, and says each will be described in detail in a forthcoming technical report:

• Decoupled Horizon Forecasting (DHF). Causal decoders trained with contiguous patch masking, as in Toto 2.0 and TimesFM-3, mask out the values to be forecast, so those values are lost as context for later forecasts. DHF keeps the full history in the sequence, and learned forecast tokens read the history before them. Every observed value serves as both target and context, and true gaps stay flagged as missing.
• Covariate-aware attention. Each input series is assigned a role: target variate, past-only covariate, or known-future covariate. Information flows only from covariates to targets, so a holiday calendar or weather forecast can be used without leaking a target's own future into its forecast.
• Synthetic-first data. Most training data is synthetic. The team first generated univariate processes with trends, seasonality, and diverse noise, then used a multivariate synthesizer to add nonlinear dependencies between them.
• Two-stage training. Timestral is pre-trained at scale on purely synthetic data with a context size of 8192, then post-trained on a mix of real and synthetic data with a 4x longer context. The objective is a pinball loss over 21 quantiles, a discrete approximation of CRPS.

Adapting to production data

Public benchmarks measure general performance, but production data can differ significantly in behavior and structure, and it can shift as markets change, equipment degrades, and demand patterns move. The article argues that the lasting value lies in adapting a model to in-domain data and retraining it as the distribution changes. Through Mistral Forge, the same training pipeline used for Mistral's language models, organisations can get zero-shot inference out of the box, add architectural extensions such as domain-specific covariates, fine-tune on proprietary data while preserving general performance, and control the training data, including the training mix, replayed pre-training data, and synthetic samples reflecting the target domain. The article says this control ensures nothing backtested was seen during pre-training. The full pipeline can be deployed on client hardware, including air-gapped environments.

Key Features at a Glance

• Benchmark leadership: First on GIFT-Eval, TIME, and fev-bench with a 209M model – ranks first across all five metrics, from a single checkpoint without ensembling.
• Model sizes: 5M, 28M, 123M, and 209M parameters – one architecture and training recipe, so teams can trade accuracy for speed.
• Fast CPU inference: 10 ms (5M) to 140 ms (209M) per one-day-ahead forecast on an 8-thread server CPU – no GPU needed at inference.
• Probabilistic output: 21 quantiles – shows forecast uncertainty for alert thresholds and risk-based maintenance planning.
• Covariate support: Past-only and known-future covariates, plus categorical inputs – uses schedules, weather, and machine states without leaking future target values.
• Context length: 32,768 time steps – supports long histories in a single forecast.
• Deployment options: On-prem, VPC, or air-gapped – fits regulated and sensitive industrial environments.
• Open access: GitHub repository and a modified MIT licence – lets teams run the model on their own data.

Availability

The article says Timestral is being released today, but it gives no specific release date. Timestral is available through a GitHub repository for running it on your own data, and through Mistral Forge for fine-tuning and customization. The licence is a modified MIT licence. A technical report with full results will follow. The article invites readers with more specialised forecasting problems, or interest in research collaboration, to get in touch.

The published comparison includes TimesFM-3, Toto-2.0-2.5B, Chronos-2 and t0-gamma. Lower scores are better for the GIFT-Eval and TIME error metrics, while fev-bench reports skill scores.

Benchmark Comparison

The published comparison includes TimesFM-3, Toto-2.0-2.5B, Chronos-2 and t0-gamma. Lower scores are better for the GIFT-Eval and TIME error metrics, while fev-bench reports skill scores.