Mistral's 209M-parameter Timestral tops GIFT-Eval, TIME and fev-bench time-series leaderboards

Mistral released Timestral on October 15, 2026, its first family of time-series foundation models. According to Mistral, the largest model has 209M parameters and ranks first on the three main public forecasting benchmarks, GIFT-Eval, TIME and fev-bench, beating the previous state of the art with about two-thirds of its parameter count.

What was released

Timestral comes in four sizes: 5M, 28M, 123M and 209M parameters. All share one architecture and training recipe. The models produce probabilistic (quantile) multivariate forecasts. They support past-only covariates, known-future covariates and categorical covariates directly, without numeric encoding.

Mistral says fine-tuning for domain adaptation is available through Mistral Forge. Pricing, context window length and weight availability were not disclosed in the source material reviewed.

Benchmark claims

Mistral says that among pretrained, zero-shot models, Timestral ranks first across all five metrics on the three benchmarks, using a single checkpoint without ensembling:

• GIFT-Eval (GMean MASE and GMean WQL): ranked against 83 comparable pretrained and zero-shot entries.
• TIME (GMean MASE and GMean CRPS): 35 leaderboard entries.
• fev-bench (SQL skill score): 37 leaderboard entries.

Mistral says results have been submitted to the GIFT-Eval and TIME leaderboards. Exact scores were presented in charts and are not reproduced in the text, so they cannot be quoted here.

Mistral also claims every Timestral size sits on the Pareto front of accuracy versus model size. In its comparisons, the 28M model beats Chronos-2, which it describes as about four times larger, on GIFT-Eval and TIME. The 209M model beats TimesFM-3 and Falcon-2.0 with fewer parameters. These comparisons are Mistral's own.

CPU inference

Timestral does not require a GPU at inference. On a standard server CPU with 8 threads, forecasting one day ahead from three weeks of 15-minute history takes:

• 10 ms with the 5M model
• 140 ms with the 209M model

The 28M model is five times faster than the 209M model. The 123M model comes within 0.2% of the 209M model's score at three quarters of its latency, according to Mistral.

Technical approach

Mistral lists four contributions, to be detailed in a forthcoming technical report:

• Decoupled Horizon Forecasting (DHF): Unlike causal decoders trained with contiguous patch masking (as in Toto 2.0 and TimesFM-3), DHF keeps the full history in the sequence. Learned forecast tokens read the history before them, so every observed value serves as both target and context. True gaps stay flagged as missing.
• Covariate-aware attention: Each series is assigned a role (target, past-only covariate or known-future covariate). Information flows only from covariates to targets, which prevents leakage of a target's own future.
• Synthetic-first data: Most training data was synthetic. Mistral generated univariate processes with trends, seasonality and varied noise, then linked them with a multivariate synthesizer that adds nonlinear dependencies.
• Two-stage training: Pre-training at scale followed by a second stage. The source text was cut off before the details.

What this means

Timestral takes Mistral beyond language and multimodal models into numerical forecasting, where the market is enterprise telemetry, observability and industrial sensor data. The pitch is efficiency. A 5M–209M model that runs in 10–140 ms on CPU can sit inside existing monitoring pipelines, where larger forecasting models may be impractical to deploy.

The rankings are Mistral's own reading of public leaderboards, and the first-place claim covers pretrained zero-shot models only. Independent verification will depend on leaderboard acceptance and the technical report. The missing licensing, weight-access and pricing details will determine how widely Timestral can be adopted outside Mistral's platform. Mistral's pairing of Timestral with language-model agents for root-cause analysis suggests it intends to sell forecasting as one component of agentic enterprise systems rather than as a standalone product.