Mistral's Timestral: A 209M Time-Series Model That Beats the Giants on CPU

Mistral's Timestral: A Tiny Time-Series Model That Beats the Giants

Mistral AI has entered the time-series forecasting arena with a bold statement. On October 15, 2026, the French AI lab released Timestral, its first family of foundation models designed specifically for time-series data. The headline number is striking: the largest Timestral model, with just 209 million parameters, claims the top spot on all three major public forecasting benchmarks—GIFT-Eval, TIME, and fev-bench—while running inference entirely on a standard server CPU.

This is a significant move for a company primarily known for large language models like Mistral Large. By focusing on efficiency and accessibility, Mistral is challenging the assumption that state-of-the-art forecasting requires massive, GPU-hungry models. The release positions Timestral as a practical tool for enterprises that deal with sensor telemetry, energy load, supply chain demand, and financial volatility.

What Is a Time-Series Foundation Model?

Time-series data is everywhere in enterprise: equipment sensors, observability metrics, energy consumption, sales figures, and market prices. Traditional forecasting methods like ARIMA and linear regression are well-established but limited—they are typically fitted to a single series and learn only from that series' history. Deep learning models like PatchTST can learn nonlinear patterns across related series, but they are usually trained from scratch on each dataset, meaning their knowledge doesn't transfer to new domains.

Time-series foundation models (TSFMs) bring the pre-training paradigm to this field. Trained on a large and diverse corpus, a TSFM learns patterns that transfer across series, allowing it to forecast zero-shot on data it has never seen. This is a paradigm shift: instead of building a custom model for every data stream, you can use one pre-trained model that works out of the box, even in domains where data is scarce.

Benchmark Dominance with a Fraction of the Parameters

Mistral's claims are backed by results on the three main public forecasting benchmarks. According to the company, Timestral ranks first among pre-trained, zero-shot models across all five metrics on GIFT-Eval, TIME, and fev-bench, using a single checkpoint without ensembling.

On GIFT-Eval, the flagship 209M model achieves a MASE of 0.658, narrowly edging out t0-gamma (0.661). On fev-bench, it scores a SQL skill score of 50.39, ahead of Google's TimesFM-3 at 48.99. The efficiency story is even more compelling at smaller scales: the 28M model beats Chronos-2—which is about four times larger—on both GIFT-Eval and TIME. The 209M model outperforms TimesFM-3 and Falcon-2.0 with fewer parameters, placing the entire family on the Pareto front of accuracy versus model size.

These results are particularly notable because they come from a model that doesn't require a GPU. On a standard server CPU with 8 threads, forecasting one day ahead from three weeks of 15-minute history takes just 10 ms with the 5M model and 140 ms with the 209M model. This opens the door to real-time, on-premise forecasting in environments where GPU resources are unavailable or cost-prohibitive.

Key Technical Innovations

Timestral's performance is not accidental. Mistral's technical report (forthcoming) describes several key contributions that combine to produce these results.

Decoupled Horizon Forecasting (DHF): Traditional causal decoders trained with contiguous patch masking, as seen in Toto 2.0 and TimesFM-3, mask out the values to be forecast. This makes them look like missing values and also loses them as context for later forecasts. DHF separates these two roles: the full history stays in the sequence, and learned forecast tokens read the history before them. This means every observed value serves as both target and context, true gaps stay flagged as missing, and training horizon lengths can be tuned independently of missing-data augmentation.

Covariate-aware attention: Timestral assigns a role to each input series: target variate, past-only covariate, or known-future covariate. Known-future covariates (like a holiday calendar or weather forecast) are visible across the whole window, while targets remain causal. Information flows only from covariates to targets, preventing the model from leaking a target's own future into its forecast. This is crucial for practical applications where production schedules or planned shutdowns are known in advance.

Synthetic-first data: The majority of Timestral's training data was synthetic. Mistral first generated univariate processes with coherent temporal dynamics—trends, seasonality, and diverse noise patterns—then used a multivariate synthesizer to create complex, nonlinear dependencies between these processes. This approach allows the model to learn fundamental time-series patterns without being limited by the biases and scarcity of real-world datasets.

Two-stage training: Timestral is first pre-trained at scale on purely synthetic data with a context size of 8,192, then post-trained on a mix of real and synthetic data with a 4x longer context (32,768 time steps). The objective is a pinball loss over 21 quantiles, a discrete approximation of CRPS, which enables probabilistic forecasting.

Practical Capabilities for Enterprise

Timestral is not just a benchmark toy; it's designed for real-world deployment. The model supports multivariate forecasting, meaning it can model related series jointly—a pump's vibration, temperature, and pressure are not independent, and Timestral can exploit those dependencies. It also provides probabilistic forecasts, predicting a full set of quantiles so teams can see the uncertainty range and set alert thresholds that reflect risk, rather than relying on a single point estimate.

One of the most distinctive features is support for categorical covariates. Things like "day of the week" or "machine state" (running, idle, under maintenance) don't have a natural numerical order. Encoding them as numbers would impose a false scale. Timestral takes them as categorical inputs directly, improving forecast quality.

Finally, Timestral can condition on known-future covariates—production schedules, planned shutdowns, or weather forecasts—so a scheduled load increase is expected rather than flagged as an anomaly. This makes the model far more useful in operational settings where future events are known but their impact on the target series is not.

From Model to System: Agentic Workflows

Mistral envisions Timestral as more than a standalone forecasting tool. It can run inside existing monitoring pipelines or be paired with language models in multi-modal agentic workflows. In an industrial plant, for example, Timestral could run continuously on sensor streams, detecting anomalies as they occur. When it flags an anomaly on a pump, an agent could embed the signals and compare them against historical failure data, then spawn sub-agents to perform root cause analysis using maintenance records, shift logs, equipment manuals, and past work orders. The result is a diagnosis and a plan backed by sensor data, rather than just an alert.

Adaptability and Open Weights

Public benchmarks measure general performance, but production data can be significantly different and may shift over time. Timestral is available through Mistral Forge, the same training pipeline used for Mistral's language models. This supports zero-shot inference, architectural extensions (like domain-specific covariates), fine-tuning on proprietary data, and full control over the training data mix. The entire pipeline can be deployed on clients' hardware, including air-gapped environments—a key selling point for enterprises with strict data sovereignty requirements.

The model is released under a Modified MIT license, and the weights and inference code are available on GitHub and Hugging Face. The 209M model's weights are approximately 837 MB in fp32, making it easy to distribute and run.

Market Context and What's Next

Timestral's release is a direct challenge to Google's TimesFM-3, which was previously one of the most visible entries in the TSFM category. The near-miss on GIFT-Eval and the clear win on fev-bench will likely prompt a competitive response from Google. For Mistral, this is a strategic move into a part of the AI market that gets far less attention than chatbots and coding assistants, but which has enormous enterprise value.

The company has published a GitHub repository so developers can run Timestral on their own data immediately, with a full technical report to follow. For organizations with specialized forecasting problems, Mistral is open to collaboration. The combination of state-of-the-art accuracy, CPU-only inference, and open weights makes Timestral a compelling option for any team looking to modernize their forecasting stack.

As time-series data continues to grow in volume and importance, foundation models like Timestral could become as essential as the LLMs that dominate the headlines. The promise is simple: better forecasts, less compute, and full control over your data.