Timestral Guide: Mistral’s No-Training Forecasting AI — Predict Sales in 10ms on a Plain CPU

Timestral is an AI that predicts the future from history — no training required. Think of it like a weather forecast, but for your sales numbers: you hand it your past data, and it tells you what’s likely to come next — with an honest “I’m this sure” range attached. Launched October 15, 2026 by Mistral AI, it’s Mistral’s first “time-series foundation model” family, in four sizes (5M, 28M, 123M, 209M parameters). Time series just means data measured over time — daily sales, website visits, electricity use. You hand it your past numbers; it returns future values with confidence intervals (its honest “I’m this sure” range), in 10 milliseconds, on an ordinary CPU. No GPU, no training job, no data-science degree.

Why this matters to you

Forecasting used to be a specialist service: collect data, train a model, maintain it. Timestral’s zero-shot mode — prediction without training — deletes most of that work. This is what it means for you: if a small shop asks “how much stock should I order for next month,” you can now answer with an actual AI forecast instead of a spreadsheet guess, and the compute cost is essentially zero.

The four sizes and the price

The official repository is on Hugging Face: mistralai/Timestral-209M (verified live). The four sizes trade accuracy against speed:

Size	Best for	Notes
5M	Edge devices, cheap VPS	~10ms forecast on an 8-core server CPU, per Mistral’s official release
28M	Most small-business jobs	Sweet spot: better accuracy, still fast
123M	Noisy, complex series	More capacity for hard patterns
209M	Maximum accuracy	Heaviest, still small by 2026 standards

Price: the models are downloadable. There’s no per-call API price — you run them yourself. Code is Apache 2.0 (free for almost everything). The catch: weights use a Modified MIT license — companies with over $20M in annual revenue need a separate authorization from Mistral (see pitfalls).

The money play: forecasting as a service for small merchants

Who buys: Shopify store owners, restaurant chains, subscription-box sellers — anyone who wastes money on wrong inventory. What you sell: a weekly “what will sell next month” forecast report, or a small dashboard fed by their sales data. What you charge: $200–$400/month per client; it costs you pennies in compute. 7-day validation: Ask one merchant for 6–12 months of daily sales. Run the five steps above on days 1–2. Show them the forecast-vs-reality chart on day 3. If the chart is right more often than their gut, you have a client; sell a 3-month pilot.

Run your first forecast in five steps

Ready to run your first forecast? Here’s the five-step path.

1. Get a Linux server or laptop with Python. Any 8-core machine works — the 5M model needs no GPU.
2. Download the weights. Open huggingface.co/mistralai/Timestral-209M, accept the license terms, and download the model files (start with the 5M or 28M size for speed).
3. Prepare your history. Put your past data in a CSV file — one column of dates, one column of values (e.g., daily revenue for the last 90 days). More history, better forecast.
4. Run a zero-shot prediction. Feed the history to the model and ask for, say, the next 30 values. You get 21 quantiles back — think of them as confidence bands: the middle band is the most likely path, the outer bands show the optimistic and pessimistic range.
5. Sanity-check against reality. Hold out your last 30 days, forecast them, and compare. If the middle band tracks reality, you have a service; if not, try a larger size or a longer history.

What a round trip looks like (illustrative example):

Your input — a plain CSV of daily revenue:

date,revenue
2026-07-24,1280
2026-07-25,1315
2026-07-26,1290

What comes back — 21 numbers per future day, forming three bands (illustrative):

day +1:  low 1245 … most-likely 1335 … high 1420
day +2:  low 1220 … most-likely 1310 … high 1445
…
day +30: low 1180 … most-likely 1360 … high 1520

The middle line is the model’s best guess; the spread is how sure it is. A narrow band means “history speaks clearly here”; a wide one means “expect surprises.”

Pitfalls and limitations

• Not plug-and-play. Honest warning: this needs some technical comfort — Python, CSVs, a server. It’s not a website you click through. Budget learning time or partner with someone technical.
• License ceiling. Code is Apache 2.0, but weights are Modified MIT: companies with annual revenue above $20M need separate authorization from Mistral. Small-business clients are fine; don’t resell the model itself to enterprises without checking the official terms.
• Forecasts are probabilistic, not prophecy. The 21 quantiles are the point — read the confidence bands, not just the middle line. Black-swan events (a viral TikTok, a supply shock) are outside any model’s history.
• Garbage in, garbage out. Forecasts only know what history taught them. Dirty or thin data (under ~60 data points) gives shaky results.