Mistral AI's official family of time-series foundation models (5M to 209M parameters) delivering state-of-the-art probabilistic forecasting on ordinary CPUs without needing a GPU.
| Step 1: Input Telemetry | ➔ | Step 2: Zero-Shot Attention | ➔ | Step 3: Quantile Deliverables |
|---|---|---|---|---|
| Ingest past series & future calendar covariates up to 32,768 steps | ➔ | Decoupled horizon tokens process sequence on standard CPU in 10-140ms | ➔ | Receive 21 probabilistic quantiles (P01 to P99) with risk thresholds |
Install the official timestral library from Mistral AI's GitHub repository along with dependencies, and execute the production pipeline:
pip install git+https://github.com/mistralai/timestral.git yfinance matplotlib torch; python main.pyOr install in your active virtual environment:
pip install git+https://github.com/mistralai/timestral.git
pip install yfinance matplotlib torch
python main.pyTimestral-209M ranks #1 globally across all three major public time-series benchmarks, outperforming models more than 10x its size:
| Model | Parameters | GIFT-Eval MASE (↓) | TIME MASE (↓) | fev-bench SQL (↑) | Inference Device |
|---|---|---|---|---|---|
| Timestral-209M | 209M | 0.658 | 0.626 | 50.39 | Standard CPU (140ms) |
| t0-gamma | ~250M | 0.661 | 0.637 | 49.42 | GPU Recommended |
| TimesFM-3 (Google) | 330M | 0.667 | 0.640 | 48.99 | GPU Recommended |
| Toto-2.0-2.5B (Databricks) | 2,500M | 0.696 | 0.642 | 44.42 | Heavy GPU Required |
| Chronos-2 (Amazon) | 120M | 0.698 | 0.662 | 47.28 | GPU Standard |
-
Stock Price Trend & Risk Tunnel (
outputs/project_1_stock_trend_forecaster.png):- Automatically loads historical price data (e.g. NVDA) via
yfinance. - Forecasts the upcoming 21 trading days with macro momentum curve and dual-layer P10-P90 uncertainty tunnels.
- Automatically loads historical price data (e.g. NVDA) via
-
Store Sales & Stock Outage Planner (
outputs/project_2_store_sales_planner.png):- Forecasts daily retail item demand conditioned on upcoming discount promotions using
future_covariates. - Generates confidence bands to guarantee shelves never run out of inventory.
- Forecasts daily retail item demand conditioned on upcoming discount promotions using
-
Business Cash Flow & Safety Balance (
outputs/project_3_cash_flow_safety.png):- Projects weekly business cash balances and incoming revenue trends with quarterly billing oscillations.
- Highlights worst-case P10 cash dips against critical emergency cash reserve lines.
- Smart IoT & Edge Telemetry: Run the 5M model on low-power devices for instant anomaly alerts in 10ms.
- Renewable Energy Grid Balancing: Predict power demand using known future weather forecasts as covariates.
- Supply Chain Lead Time Planning: Forecast warehouse inventory bottlenecks weeks in advance.
- Financial Treasury Reserves: Project cash inflows and outflows across multiple bank accounts.
- SRE & Cloud Cost Optimization: Auto-scale cloud server instances based on predicted usage curves.
- Automated LLM Incident Reports: Pairing numerical forecasts with Mistral Large for written executive summaries.
- Streaming Sliding-Window Mode: Real-time socket ingestion for sub-second live sensor feeds.
- ONNX & OpenVINO Quantization: Running Timestral on microcontrollers and embedded NPU chips.
- Cross-Correlated Market Pairs: Joint multivariate spread forecasting for algorithmic market making.
- Zero-Code Dashboard UI: Web-based interactive drag-and-drop CSV forecaster with interactive charts.
Built upon Mistral AI's official open-weight release. Released under Modified MIT license. Official repository: github.com/mistralai/timestral.
mistral timestral, timestral 209m, timestral 5m, timestral 28m, timestral 123m, mistral ai, timesfm 3.0, amazon chronos, gift eval, time benchmark, fev bench, zero shot time series, forecasting foundation model, cpu ai forecasting, no gpu ai, stock prediction ai, demand forecasting, multivariate forecasting, local ai, python time series, yfinance python