GitHub radar
Forecast your own number series without training a model
Timestral is a small Mistral model that predicts future values for any numeric series, like daily sales, without needing training on your data.
It's a ready-made forecasting model: feed it a history of numbers, and it returns a forecast with a range of likely future values, not just one number.
When it helps
Useful for a quick read on trends in sales, signups, or other numeric metrics, run locally and for free. For high-stakes financial decisions, use dedicated analytics tools instead.
Pros
- Runs locally on a regular laptop, no GPU needed
- No training on your own data required
- Gives a range of outcomes, not just one number
- Free, no subscriptions or API keys
Cons
- Not a replacement for serious analytics on big-money decisions
- Needs a coding agent to run, no ready-made interface
How to set it up — step by step
- 1Open a coding agent (Claude Code, Codex) in your project folder.
- 2Send the agent the text from the block below.
- 3Prepare your numeric series — for example, daily sales over recent weeks.
- 4Check that the agent installed the model from https://github.com/mistralai/timestral/ and ran it on your laptop.
- 5Look at the forecast range and compare it to real data later.
Text for your agent
Copy this and send it to your agent — Claude Code, Codex, any of them:
Set up Mistral's time-series forecasting model Timestral-209M for me: https://huggingface.co/mistralai/Timestral-209M — run it locally through their GitHub repository, show me how to feed it a column of my own numbers (for example daily sales), and print a forecast with a range for the next few weeks.
A regular laptop - the model is tiny, under one gigabyte, and runs even without a graphics card, on the CPU alone.
Open on Hugging Face▌ More finds