How to Run Timestral-209M Locally: Mistral's 209M-Parameter Time Series Forecasting Model

Timestral-209M is a 209 million parameter time series foundation model from Mistral AI. It runs zero-shot forecasting on CPU or GPU, accepts up to 32,768 historical time steps, and outputs 21 quantiles per prediction. It tops all three major zero-shot forecasting benchmarks as of October 2026.

Written by Priya Raghunathan, local-AI hardware reviewer. I verify parameter counts, benchmark numbers and install commands against primary sources before recommending a setup path.

A new top-20 entrant on Hugging Face's trending list

Timestral-209M entered the top 20 of Hugging Face's overall trending list in October 2026. At time of writing it has 219 likes and 4,008 downloads on the mistralai/Timestral-209M model page, released October 14, 2026. The GitHub repo has 640 stars.

Specs at a glance

What it actually does

Timestral is a family of time series foundation models built for zero-shot forecasting. That means you load the pretrained checkpoint and run predictions on your data without any fine-tuning step. The 209M version is the largest in the family, which also ships 5M, 28M and 123M checkpoints. It handles both univariate forecasting (a single series) and multivariate forecasting (multiple series with shared structure), supports past-only and known-future covariates, and marks missing values as NaN rather than requiring you to impute them before input.

The architecture divides a time series into patches of 32 steps each, then runs those patches through 24 transformer layers with separate time attention and variate attention mechanisms. Predictions come back as 21 quantiles, from 0.01 to 0.99, which gives you a full uncertainty distribution rather than a single point estimate. Longer forecasts beyond 1,024 steps work block by block, with each block's median appended to the context before the next block runs.

Benchmark performance

Timestral-209M tops three major zero-shot forecasting benchmarks. For MASE and CRPS, lower is better. For fev-bench skill scores, higher is better.

How the family scales

Performance increases consistently from 5M to 209M parameters, though the jump from 123M to 209M is incremental rather than dramatic.

Install it

Timestral ships from GitHub. You need Python 3.10 or newer and torch 2.4 or newer. Install directly with pip:

Or clone the repo and use uv to lock to the exact versions Mistral tested against:

Run a forecast in Python

The example below uses synthetic data. Swap in a real NumPy array of historical values from your own dataset.

import numpy as np
import torch
from timestral import TimestralForecaster

model = TimestralForecaster.from_pretrained(
    "mistralai/Timestral-209M",
    device="cuda" if torch.cuda.is_available() else "cpu"
)

# 1,000 historical observations
context = np.random.default_rng(42).standard_normal(1000).cumsum()
forecast = model.forecast(context, prediction_length=96)

median = forecast.median
lower, upper = forecast.quantile(0.1), forecast.quantile(0.9)

The model auto-selects CUDA if a GPU is available and falls back to CPU otherwise. The forecast object returns a median and the full set of 21 quantiles, so you can extract any confidence interval you need.

License

FAQ

What is Timestral-209M?

Timestral-209M is a 209 million parameter time series forecasting model from Mistral AI. It runs zero-shot on CPU or GPU, accepts up to 32,768 historical time steps, and outputs probabilistic forecasts as 21 quantiles.

Do I need a GPU to run it?

No. The model card specifies PyTorch on CPU or GPU. A GPU will be faster, but the model loads and runs on CPU. No specific VRAM requirement is documented.

Does it need fine-tuning?

No. Timestral-209M is a foundation model designed for zero-shot use. You load the pretrained checkpoint and run it on your data directly.

How does it handle missing values?

Missing values are passed as NaN. The model embeds each patch with a missing-value indicator alongside the raw values, so you do not need to impute before inference.

What is the difference between the 5M and 209M versions?

The 209M version scores best on all three benchmarks, but the 5M version is much lighter on hardware and still useful for many practical series. Try the smallest checkpoint that meets your accuracy requirement rather than defaulting to the largest.

Where to go from here

• Browse other local AI models by category: the local AI model directory lets you compare specs side by side.