Mistral releases Timestral-209M, a time series model claiming top zero-shot scores on three benchmarks

Mistral AI has released Timestral-209M, a 209-million-parameter time series foundation model for zero-shot forecasting. It is the largest of the Timestral family, which also includes 5M, 28M and 123M checkpoints. According to the model card, it posts the best scores on three public forecasting benchmarks: GIFT-Eval, TIME and fev-bench.

Specifications

• Parameters: 209M (F32 Safetensors)
• Architecture: Patch-based transformer with time attention and variate attention; 24 layers, embedding dimension 768
• Patching: 32 time steps per patch; each patch embeds its values together with missing-value indicators
• Context length: Up to 32,768 time steps (1,024 patches)
• Forecast horizon: Up to 1,024 time steps (32 patches) in one forward pass. Longer horizons are produced block by block, with each block's median forecast appended to the context
• Outputs: 21 quantiles, from 0.01 to 0.99
• Inputs: Univariate or multivariate targets, plus past-only and known-future covariates, numeric or categorical. Categorical covariates are target-encoded, and missing values are marked as NaN
• Inference: PyTorch on CPU or GPU (Python 3.10+, torch 2.4+)

Context length is measured in time steps, not text tokens, so it is not comparable to LLM context windows.

Benchmark results

The scores below come from the model card, which links to the GIFT-Eval, TIME and fev-bench leaderboards. Lower is better for MASE and CRPS; higher is better for skill scores.

Benchmark	Metric	Timestral-209M	t0-gamma	TimesFM-3	Toto-2.0-2.5B	Chronos-2
GIFT-Eval	MASE ↓	0.658	0.661	0.667	0.696	0.698
GIFT-Eval	CRPS ↓	0.446	0.452	0.456	0.476	0.485
TIME	MASE ↓	0.626	0.637	0.640	0.642	0.662
TIME	CRPS ↓	0.522	0.534	0.536	0.539	0.556
fev-bench	SQL skill ↑	50.39	49.42	48.99	44.42	47.28
fev-bench	MASE skill ↑	39.49	37.99	37.82	32.54	35.50

The margins over the closest competitor, t0-gamma, are small on GIFT-Eval (0.658 vs 0.661 MASE) and larger on TIME (0.626 vs 0.637 MASE) and fev-bench (50.39 vs 49.42 SQL skill score).

Scaling across the family

Gains flatten at the top of the range. On GIFT-Eval MASE, the four sizes score 0.705 (5M), 0.678 (28M), 0.659 (123M) and 0.658 (209M). On fev-bench SQL skill score they reach 46.19, 48.42, 50.18 and 50.39. By the card's own table, the 123M model already beats every listed competitor on both GIFT-Eval and TIME MASE and on fev-bench SQL skill score.

Availability and license

Weights are on Hugging Face, with source code and benchmark reproduction scripts in Mistral's GitHub repository. The license is a Modified MIT license that bars uses that infringe or violate third-party rights, including intellectual property rights. No Inference Provider currently hosts the model, and no hosted API pricing has been disclosed. A technical report is listed as "coming soon," and a blog post accompanies the release. The model page shows 2,601 downloads in the last month.

What this means

This is a notable move by Mistral into a domain far from its usual LLM work. Time series foundation models are a crowded field, and the benchmark gaps here are narrow. Without the tech report, the training data, the evaluation protocol and any contamination checks are not yet independently verifiable. Leaderboard leads of a few thousandths of MASE should be treated as parity until others reproduce them.

The more practical result is efficiency. Because the 123M and 209M checkpoints score nearly the same, and the 28M model beats two of the four listed competitors on GIFT-Eval, teams can likely run smaller checkpoints on CPU with limited accuracy loss. The listed Toto-2.0-2.5B, whose name suggests 2.5B parameters, is roughly 12x larger than Timestral-209M and scores lower in every row of the table.

Support for covariates, multivariate inputs and 21-quantile output makes the model usable for demand planning, energy and capacity forecasting, where calibrated uncertainty matters as much as point accuracy. Real-world performance on private datasets remains the open question.