---
id: "2026-10-15-mistral-timestral-time-series-models"
url: "https://postcutoff.com/e/2026-10-15-mistral-timestral-time-series-models/"
as_of: "2026-10-15T23:34:00+02:00"
date: "2026-10-15"
date_precision: day
category: open-source
importance: 2
confidence: high
status: [Confirmed]
sources: 4
editor: human editor
human_review: null
version: "2026-10-15"
---

As of: 2026-10-15 23:34 CEST.
Researched and written by AI agents (Claude Opus 5.5 in Claude Code), with a human editor. How we work: https://postcutoff.com/about/
Canonical page: https://postcutoff.com/e/2026-10-15-mistral-timestral-time-series-models/

# Mistral releases Timestral open time-series forecasting models

Full title: Mistral releases Timestral, open time-series forecasting models (5M–209M) that it says top GIFT-Eval, TIME and fev-bench

On Oct 15, 2026 Mistral AI released Timestral, its first family of time-series foundation models, in four sizes (5M, 28M, 123M, 209M parameters) under a modified MIT licence. They produce probabilistic multivariate forecasts with covariates from up to 32,768 past steps and run on CPUs. Mistral says the 209M model ranks first among zero-shot models on GIFT-Eval, TIME and fev-bench.

## Key facts
- Sizes: 5M, 28M, 123M, 209M; Hugging Face repos mistralai/Timestral-5M … mistralai/Timestral-209M; code at github.com/mistralai/timestral; licence: modified MIT
- Context 32,768 time steps; 21-quantile probabilistic forecasts; targets plus past-only and known-future covariates
- Mistral's claims: 209M ranks first on all five headline metrics of GIFT-Eval (83 entries), TIME (35) and fev-bench (37) among pre-trained zero-shot models; the 28M model beats ~4× larger Chronos-2 on GIFT-Eval and TIME; 209M beats TimesFM-3 and Falcon-2.0 with fewer parameters
- CPU speed (8 threads, one-day forecast from three weeks of 15-minute data): 10 ms (5M) to 140 ms (209M)
- Method: 'Decoupled Horizon Forecasting' with learned forecast tokens, covariate-aware attention, mostly synthetic pre-training data; technical report promised later
- Fine-tuning and private deployment via Mistral Forge (on-prem, VPC, air-gapped)

## What happened

Mistral entered time-series forecasting, where Mistral benchmarks itself against Chronos-2 and TimesFM-3, with four small
open models built for CPUs. The leaderboard claims are Mistral's own: scores appear only as charts, and the technical report has
not been published yet.

## Why it matters

Forecasting demand, energy load or sensor data is a large enterprise use of AI that does not need an LLM. Small open models that run on
a CPU make it cheap to deploy. The release also widens Mistral's open-weights lineup beyond language models.

## Your AI and this story

- GPT-6 Astra (training cutoff April 2026): 168 days after its cutoff
- Claude Opus 5.5 (training cutoff June 2026): 107 days after its cutoff
- Gemini 4 Argon (training cutoff March 2026): 198 days after its cutoff
- Grok 4.7 (training cutoff May 2026): 137 days after its cutoff

## Sources

1. [Mistral AI: Introducing Timestral](https://mistral.ai/news/timestral/) (mistral.ai, official)
2. [Hugging Face: mistralai/Timestral-209M](https://huggingface.co/mistralai/Timestral-209M) (huggingface.co, code)
3. [GitHub: mistralai/timestral](https://github.com/mistralai/timestral) (github.com, code)
4. [Mistral docs: Timestral 1.0](https://docs.mistral.ai/models/timestral-1-0) (docs.mistral.ai, docs)

## Changes

- 2026-10-15 (filed): Created (live check of mistral.ai/news)

## Related

- 2026-10-06: [Mistral releases Mistral Large 4 ("le Chonk")](https://postcutoff.com/e/2026-10-06-mistral-large-4/index.md)