Saját szervereken, GPU nélkül futtatható az új Timestral modellsorozat

Mistral AI has released Timestral , its first family of time series foundation models. Timestral is a set of 4 zero-shot forecasting models from Mistral AI, sized from 5M to 209M parameters. The largest checkpoint ranks first among pretrained, zero-shot models on GIFT-Eval, TIME and fev-bench. It does this without a GPU at inference. For developers, this Timestral time series foundation model means probabilistic forecasts on standard server CPUs, with open weights and inference code . TL;DR Size: 4 models: 5M, 28M, 123M, 209M parameters (dense). Context up to 32,768 time steps. Runs on: Standard server CPU, no GPU needed. 209M weights are 837 MB in fp32. GPU optional. Performance: #1 among zero-shot pretrained models on GIFT-Eval, TIME and fev-bench (Mistral-reported). Best: fev-bench SQL skill score of 50.39, ahead of TimesFM-3 at 48.99. Worst: GIFT-Eval MASE of 0.658 leads t0-gamma (0.661) by only 0.003. What is Timestral? Timestral is a time series foundation model (TSFM). It is pretrained on a large, diverse corpus, then forecasts zero-shot on data it has never seen. Classical tools like ARIMA fit one series at a time. Deep models like PatchTST usually train from scratch on each dataset. Timestral transfers patterns across domains instead. Each model outputs 21 quantiles per time step . It forecasts multivariate targets jointly. It accepts past-only, known-future and categorical covariates. How does Timestral work? Timestral is a causal, patch-based transformer with time attention and variate attention. The 209M model uses 24 layers and an embedding dimension of 768 . Each patch covers 32 time steps. A single forward pass forecasts up to 1,024 steps. Longer horizons run block by block, feeding each block’s median back as context. Mistral lists 4 technical contributions : Decoupled Horizon Forecasting (DHF): Learned forecast tokens read the full history. Observed values stay in context instead of being masked, unlike contiguous patch masking in Toto 2.0 and TimesFM-3. Covariate-aware attention: Each series is a target, past-only covariate 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. The loss is pinball loss over 21 quantiles. How does Timestral perform on benchmarks? All scores below are zero-shot, from a single checkpoint without ensembling. They come from the Timestral-209M model card . GIFT-Eval: MASE 0.658, CRPS 0.446 (lower is better). TIME: MASE 0.626, CRPS 0.522 (lower is better). fev-bench: SQL skill score 50.39, MASE skill score 39.49 (higher is better). The size scaling is flat at the top. Timestral-123M scores 0.659 MASE on GIFT-Eval , against 0.658 for 209M. Mistral team states that the 28M model beats Chronos-2 on GIFT-Eval and TIME. Chronos-2 is about 4 times larger. How fast is Timestral on CPU? Mistral team tested a standard server CPU with 8 threads. The task: forecast 1 day ahead from 3 weeks of 15-minute history. The 5M model took 10 ms and the 209M model took 140 ms . The 28M model is 5 times faster than 209M. The 123M model lands within 0.2% of the 209M score at three quarters of its latency. How does Timestral compare with TimesFM-3, Chronos-2 and Toto 2.0? Feature Timestral-209M TimesFM-3 Chronos-2 Toto-2.0-2.5B Developer Mistral AI Google Research Amazon Datadog Parameters 209M 330M 120M 2.5B Architecture Causal patch transformer, DHF Decoder-only, contiguous patch masking Encoder-only, group attention Decoder-only, contiguous patch masking Max context 32,768 steps Not disclosed 8,192 steps Not disclosed Single-pass horizon 1,024 steps Not disclosed 1,024 steps (max) Not disclosed Quantiles 21 9 Not disclosed 9 Weights licence Modified