huggingface.co web signal

Mistral Opens Timestral Family From 5M to 209M Parameters

4 sources tracking this story
Mistral Open Source ai-models

TL;DR

  • The 28M Timestral model outperforms Chronos-2 despite being 4x smaller, undercutting the assumption that larger pretrained models dominate zero-shot forecasting.
  • CPU-only inference at 10ms (5M model) to 113ms (209M) on an 8-thread server removes GPU as a deployment requirement for edge and embedded use cases.
  • Benchmark leadership across GIFT-Eval, TIME, and fev-bench was achieved without ensembling, suggesting genuine architectural gains from Decoupled Horizon Forecasting.

Mistral has released Timestral as a family of four probabilistic time-series forecasters at 5M, 28M, 123M, and 209M parameters, with the two largest sitting first and second on every reported metric across GIFT-Eval, TIME, and fev-bench. All four lie on the Pareto front.

The design choice doing the work is Decoupled Horizon Forecasting. In the paper's own words, DHF "separates prediction tokens from the causally encoded context, so that longer horizons are trained without masking context and each observed value is both a forecast target and context for later forecasts." The alternative recipe, consecutive-patch masking used in TiRex, Toto-2.0, and TimesFM-3, couples horizon distribution to context masking, so training longer forecast distances removes observations from the context of later forecasts in the same sequence.

The small-model results are the striking part. The paper reports that "Timestral-28M outperforms Chronos-2 and Toto-2.0-2.5B, about 4 and 90 times larger, on all six metrics," and on fev-bench Timestral-5M beats every Toto-2.0 model, including the 2.5B variant with "about 480 times as many parameters."

Inference needs no GPU. A 96-step forecast from 2,048 steps of context takes "17 to 113 ms on 8 CPU threads" on an AMD EPYC 9454, with bfloat16 outputs differing from float32 by less than 0.2%. Both Timestral-123M and Timestral-209M are smaller than t0-gamma (256M) and TimesFM-3 (331M), and Timestral-123M already ranks second on every one of the six metrics, ahead of every model outside the family.

Weights are published on Hugging Face and inference code on GitHub. The training recipe pretrains exclusively on a synthetic corpus at an 8,192-step context before post-training on a mix of real and synthetic data at 32,768. Our October 19 note covered the 209M checkpoint when it first drew attention; the full technical report now accompanies all four sizes.

What others are reporting

Coverage cluster as of 24h after publish

  1. Mistral AI Read →

    First-party launch post detailing Decoupled Horizon Forecasting architecture, two-stage synthetic-then-real training, and Mistral Forge fine-tuning with air-gapped deployment.

    Ranks first on all three of the field's main public benchmarks...outperforming the previous state of the art with only two-thirds of the model parameters.
  2. Open Data Science Read →

    Adds direct head-to-head framing of 28M Timestral vs Chronos-2 and cautions benchmark leadership does not guarantee production performance.

    The 209M model ranks first among pretrained zero-shot forecasting models across GIFT-Eval, TIME, and fev-bench.
  3. Laurence Moroney (The AI Guy) Read →

    Practitioner walkthrough with code and consumer hardware testing; notes third-party leaderboard confirmation was still pending at publish time.

    You can hand it a series it has never seen and get a forecast straight away, with no training on your data first.