Mistral's Timestral-209M Tops All Three Time-Series Benchmarks on CPU Without a GPU
Mistral released Timestral on October 15, its first family of time-series foundation models. The family runs at four sizes — 5M, 28M, 123M, and 209M parameters — and the largest checkpoint ranks first among pretrained, zero-shot models on all three of the field’s main public benchmarks, from a single checkpoint without ensembling.
Benchmark Results
All scores below are from the Timestral-209M model card and Mistral’s announcement. Lower MASE and CRPS scores are better; higher skill scores are better on fev-bench.
| Benchmark | Metric | Timestral-209M | TimesFM-3 | Chronos-2 | Toto-2.0-2.5B |
|---|---|---|---|---|---|
| GIFT-Eval | MASE | 0.658 | 0.667 | 0.698 | 0.696 |
| GIFT-Eval | CRPS | 0.446 | 0.456 | 0.485 | 0.476 |
| TIME | MASE | 0.626 | 0.640 | 0.662 | 0.642 |
| TIME | CRPS | 0.522 | 0.536 | 0.556 | 0.539 |
| fev-bench | SQL skill | 50.39 | 48.99 | 47.28 | 44.42 |
Toto-2.0 has 2.5 billion parameters and finishes behind the 209M model on every row. The 28M model, per Mistral, beats Chronos-2 on GIFT-Eval and TIME despite being roughly a quarter of its size. Results were submitted to the GIFT-Eval and TIME public leaderboards; these are vendor-reported figures pending independent verification.
CPU-First Inference
No GPU required at inference time. On a standard server CPU with 8 threads, forecasting one day ahead from three weeks of 15-minute history runs in 10ms with the 5M model and 140ms with the 209M. Mistral reports the 123M model comes within 0.2% of the 209M’s GIFT-Eval score at about three quarters of its latency. For teams forecasting thousands of series per hour, the 28M is the starting point.
The model context window is 32,768 time steps, with probabilistic output: full quantile predictions at 21 quantile levels rather than a single point estimate.
What It Actually Does
Timestral accepts multivariate targets alongside past-only and known-future covariates — holiday calendars, production schedules, planned shutdowns — as typed categorical or continuous inputs. It forecasts jointly across related series rather than treating each sensor in isolation.
Mistral’s industrial use case sketch: a plant with thousands of sensors runs Timestral continuously, an agent picks up anomalies and retrieves similar historical failures via embedding search, then drafts a diagnosis backed by maintenance records and equipment manuals. The model’s role is strictly numerical; the language reasoning layer sits above it.
Model Family and Licensing
| Size | Typical use case |
|---|---|
| 5M | High-frequency, latency-sensitive pipelines |
| 28M | Batch forecasting; beats Chronos-2 on key benchmarks |
| 123M | Within 0.2% of the 209M at 75% of its latency |
| 209M | Best accuracy, 140ms CPU inference |
License is Modified MIT. Available for zero-shot inference, fine-tuning via Mistral Forge, and air-gapped deployment. A GitHub repository ships with the release; a technical report with full methodology is forthcoming.
Competitive Context
TimesFM-3 is Google’s time-series foundation model. Chronos-2 comes from Amazon. Toto-2.0 is from Datadog. Moirai is not included in Mistral’s comparison table. Margins on GIFT-Eval MASE between Timestral and t0-gamma (0.658 vs 0.661) are small enough that production data characteristics — sampling frequency, seasonality, missing value rate — will determine which model performs better in practice. The benchmark leads over Chronos-2 and Toto-2.0 are larger and more robust.