Mistral AI releases Timestral forecasting model family, beats Google TimesFM-3
Mistral AI announced Timestral, a family of four time‑series foundation models ranging from 5 M to 209 M parameters. The models are pretrained on a large synthetic corpus and can produce probabilistic forecasts zero‑shot. The 209 M checkpoint achieved the top rank among pretrained zero‑shot models on three public benchmarks – GIFT‑Eval, TIME and fev‑bench – according to Mistral’s own evaluation.…
Key points
- Timestral family includes 5M, 28M, 123M and 209M‑parameter models; 209M tops zero‑shot scores on GIFT‑Eval, TIME and fev‑bench.
- CPU‑only inference runs 10 ms (5M) to 140 ms (209M) for a 1‑day‑ahead forecast from three weeks of 15‑minute data.
- Model weights (837 MB fp32 for 209M) and code are open, installable via pip from Mistral’s GitHub repository.
All four models run on standard server CPUs without a GPU; the 5 M model forecasts a day ahead in 10 ms, while the 209 M model needs 140 ms on an 8‑thread CPU. The models accept up to 32,768 time steps of history, output 21 quantiles per step, and support past‑only, known‑future and categorical covariates. Weights are released under a modified MIT licence and can be installed via pip from the public GitHub repository; fine‑tuning is available through Mistral Forge.
Model page: TimesFM-3 →
Mistral AI Releases Timestral: A 209M Forecasting Model Family That Outperforms Google’s TimesFM-3
MarkTechPost · 19 October 2026
Loading the full article…
This text was published by MarkTechPost and written by Asif Razzaq. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Generative AI & Models
All →- Anthropic’s Claude Opus 5.5 tops Arena AI model rankings · 3 src
- DeepSeek releases V4.1-Flash-1017 update with lower cost and higher DeepSWE score · 20 src
- KillSwitch language challenges LLMs, Claude Opus scores highest · 1 src
- The Truth Behind 'Apple and Meta Team Up on AI': It's Not Siri, It's Mac Studio | A Manager's Breakdown of Confirmed Facts, Reports, Unannounced Details, and Im · 1 src
- OpenAI makes Prompt-to-Voice generally available for developers · 3 src
Comments
via GitHub Discussions