Mistral's Timestral-209M tops zero-shot forecasting benchmarks on CPU

Mistral AI has released Timestral, a family of four time series foundation models sized from 5M to 209M parameters, and claims the largest checkpoint ranks first among pretrained, zero-shot models on GIFT-Eval, TIME and fev-bench — running on a standard server CPU with no GPU. The announcement matters because forecasting underpins inventory planning, energy load management and financial operations, a domain where per-dataset statistical tools and heavyweight neural models have so far split the market. The claims, however, are Mistral-reported; the company runs its own scorecard.

The 209M model posts a GIFT-Eval MASE (Mean Absolute Scaled Error, a normalized error measure where lower is better) of 0.658 and a fev-bench SQL skill score of 50.39, ahead of Google's TimesFM-3 at 48.99. The margin at the top is thin. Timestral leads t0-gamma (0.661) on GIFT-Eval by just 0.003 — a gap that could vanish under different data splits. The takeaway: the leaderboard position is real but narrow.

What is Timestral?

Timestral is a time series foundation model (TSFM), meaning it is pretrained on a broad corpus and then forecasts zero-shot on data it never saw during training. Classical tools like ARIMA fit one series at a time. Deep models like PatchTST typically train from scratch on each dataset. Timestral instead transfers learned patterns across domains.

Each model outputs 21 quantiles per time step, producing probabilistic rather than point forecasts. It forecasts multivariate targets jointly. It accepts past-only, known-future and categorical covariates (external variables that influence the target series). In plain words: one pretrained model replaces a fleet of per-dataset training runs.

How does the architecture work?

Timestral is a causal, patch-based transformer with two attention streams: time attention and variate attention. The 209M model uses 24 layers and an embedding dimension of 768. Each patch covers 32 time steps, and a single forward pass forecasts up to 1,024 steps. Longer horizons run block by block, feeding each block's median back as context. Context extends to 32,768 time steps.

Mistral lists four technical contributions:

• Decoupled Horizon Forecasting (DHF): learned forecast tokens read the full history; observed values stay in context rather than being masked, unlike the contiguous patch masking in Toto 2.0 and TimesFM-3.
• Covariate-aware attention: each series is typed as target, past-only or known-future covariate; information flows only from covariates to targets, preventing future leakage.
• Synthetic-first data: most training data is synthetic, with trends, seasonality, noise and nonlinear cross-series dependencies.
• Two-stage training: pretraining on synthetic data at context 8,192, then post-training on mixed data with 4x longer context, using pinball loss over 21 quantiles.

The synthetic-heavy training recipe is the caveat worth flagging: Mistral does not disclose the composition of the real-data portion, so how much the model depends on synthetic regularities remains an open question. The takeaway: the design is genuinely different from its rivals, but the training-data disclosure is thin.

How does it perform against TimesFM-3, Chronos-2 and Toto 2.0?

All scores are zero-shot, from a single checkpoint, without ensembling, per the Timestral-209M model card:

• GIFT-Eval: MASE 0.658, CRPS 0.446 — versus TimesFM-3 at 0.667/0.456, Chronos-2 at 0.698/0.485 and Toto-2.0-2.5B at 0.696/0.476.
• TIME: MASE 0.626, CRPS 0.522 — versus TimesFM-3 at 0.640/0.536 and Chronos-2 at 0.662/0.556.
• fev-bench: SQL skill score 50.39, MASE skill score 39.49 — versus TimesFM-3 at 48.99, Chronos-2 at 47.28 and Toto-2.0 at 44.42.

A caveat follows every promise: Google and Amazon do not publish matching figures on their own model pages, so the comparison table rests on Mistral's own card. Datadog's model card lists Toto-2.0 at 0.640/0.539 on TIME where Mistral's card lists 0.642/0.539 — a small discrepancy, but a discrepancy.

Size scaling is also flat at the top. Timestral-123M scores 0.659 MASE on GIFT-Eval, against 0.658 for the 209M model. Mistral states the 28M model beats Chronos-2 on GIFT-Eval and TIME despite Chronos-2 being roughly four times larger. In plain words: the smallest models deliver most of the accuracy at a fraction of the cost.

How fast is it on CPU?

Mistral tested a standard server CPU with 8 threads, forecasting one day ahead from three weeks of 15-minute history. The 5M model completed the task in 10 ms; the 209M model took 140 ms. The 28M model runs five times faster than the 209M. The 123M model lands within 0.2% of the 209M score at three quarters of its latency.

The 209M weights occupy 837 MB in fp32 — small enough for commodity hardware. GPU acceleration is optional, not required. The takeaway: the CPU claims are concrete, latency-tested numbers, and they are the release's strongest practical argument.

How can developers deploy it?

Installation runs through pip install git+https://github.com/mistralai/timestral.git, requiring Python 3.10+ and torch 2.4+. Dependencies are PyTorch, NumPy, safetensors and huggingface_hub. Weights download from Hugging Face: 21 MB for 5M, 111 MB for 28M, 490 MB for 123M. Both weights and inference code are public.

Mistral's documentation lists Timestral 1.0 209M as Public Preview. Fine-tuning on proprietary data runs through Mistral Forge, and the pipeline can deploy on-prem, in a VPC or air-gapped. The weights ship under a modified MIT licence — notably more permissive than TimesFM-3's non-commercial v1.0 licence, though Chronos-2 and Toto-2.0 use Apache 2.0. In plain words: enterprises can self-host it, but "modified MIT" deserves a read before commercial deployment.

The licensing gap, combined with CPU-only inference, is what positions Timestral against Google's offering: TimesFM-3 cannot be used commercially under its licence, and Mistral's can. Whether Mistral's benchmark lead survives independent reproduction will decide if this is a real shift in the forecasting market or another vendor-scored leaderboard win.