Models
What Is Timestral? Mistral’s Time-Series Forecasting Models, Explained
Timestral is Mistral’s first family of time-series forecasting models, released October 15, 2026. The four sizes, benchmarks, license, and how to try it.
Timestral is the newest name from Mistral, the French AI company, and it is not another chatbot. If you saw "Mistral Timestral" near the top of Hacker News this week and wondered what it actually does, the short version is this: it predicts what numbers will do next.
Below is the plain-English guide. What Timestral is, the four model sizes, what Mistral claims on benchmarks, whether it is free, and how it fits into the bigger story of AI agents.
What is Timestral?
Timestral (sometimes typed "Mistral Timestral" or "Timestral 1.0") is a set of AI models trained on time series. A time series is any list of numbers recorded over time: hourly electricity use, daily store sales, a machine’s temperature every 15 minutes. You give Timestral the history, and it forecasts the next stretch.
Mistral calls these "foundation models" because they are pre-trained once on a huge amount of data and then work on new datasets right away, with no training on your data first. People call that zero-shot forecasting. Before models like this, a company usually had to build and tune a separate forecasting model for each problem.
Mistral published the announcement as a research post on October 15, 2026, alongside a GitHub repository and a Hugging Face collection. It hit the top of Hacker News on October 19 and 20, which is when most people first saw the name.
Timestral model sizes: 5M, 28M, 123M and 209M
All four models share the same design and training recipe. The bigger ones are a bit more accurate, and the smaller ones are faster and cheaper to run. These are tiny next to chat models, which often have hundreds of billions of parameters.
| Model | Size | What Mistral says about it |
|---|---|---|
| Timestral-5M | 5 million parameters | Fastest: about 10 ms per forecast on a server CPU |
| Timestral-28M | 28 million | Beats Chronos-2, about 4 times its size, on GIFT-Eval and TIME |
| Timestral-123M | 123 million | Within 0.2% of the top model’s score at three quarters of the latency |
| Timestral-209M | 209 million | The flagship: ranked first on GIFT-Eval, TIME and fev-bench, about 140 ms per forecast |
The speed figures come from Mistral’s own test: a one-day-ahead forecast from three weeks of 15-minute data, on a standard server CPU with 8 threads. No GPU needed.
How does Timestral work?
You do not need the math to get the idea. Three things stand out in Mistral’s description:
- It gives a range, not one number. Each forecast comes as 21 quantiles, a spread from "low case" to "high case." That matters for planning: a store wants to know the chance of running out, not just the average.
- It uses side information. Mistral calls this covariate-aware attention. You can feed in extra signals, including ones known ahead of time, like a weather forecast or a holiday calendar.
- It learned mostly from made-up data. Pre-training used purely synthetic series, then a second stage mixed real and synthetic data. It reads up to 32,768 past time steps and forecasts up to 1,024 steps in one pass, per the model card.
Mistral says a full technical report will follow.
Timestral benchmarks: is it really the best?
Mistral says the 209M model ranks first among pre-trained, zero-shot models on all three main public forecasting benchmarks: GIFT-Eval, TIME and fev-bench, across all five reported metrics, from a single checkpoint. It also says the 209M model beats Google’s TimesFM-3 and Falcon-2.0 while using fewer parameters, and outperforms the previous best with about two-thirds of its size.
Two cautions. These are Mistral’s own numbers on launch day. Mistral says it submitted results to the GIFT-Eval and TIME leaderboards, so independent rankings should settle the question over the coming weeks. And benchmark wins in forecasting are often small margins, so test it on your own data before switching.
Is Timestral free and open source?
The weights are free to download from Hugging Face, released under what Mistral calls a modified MIT license. The code repository on GitHub is listed under Apache 2.0. Read the license file on the model card before using it in a commercial product, since "modified" means it is not the plain MIT license.
Mistral did not list a price for a hosted Timestral API in its announcement. For companies that want more, Mistral offers fine-tuning on private data, custom variables and private, on-premise or air-gapped setups through its Mistral Forge service.
How to try Timestral
- Open the Timestral collection on Hugging Face (start with Timestral-5M, it is the lightest).
- Follow the GitHub repo at github.com/mistralai/timestral, which runs on PyTorch on a CPU or GPU.
- Load a CSV of your own history, for example a year of daily sales, and ask for the next 30 days.
It is a developer tool. There is no Timestral chat app, and you will not find it inside a regular chatbot today.
Timestral vs TimesFM and Chronos
If you searched "time series foundation model," the names you usually meet are Google’s TimesFM and Amazon’s Chronos. Timestral is Mistral’s entry into that same race. Its pitch is size: matching or beating those models with fewer parameters, so it runs cheaply on CPUs inside a company’s own servers.
What this means for you
Timestral is a reminder that "AI" is splitting into specialists. A chat model writes, a coding model codes, and now a 209M forecasting model predicts. Mistral’s own example shows where this goes: a factory agent spots an odd sensor reading with Timestral, then digs through maintenance records and shift logs to find the cause. That is agentic AI in action, where several specialists work together on one job.
The same idea works on your phone. Agentic AI gives you a team of ready-made expert agents on iPhone and iPad, so a planning, research or writing task goes to the right expert without you choosing models. Other new specialist models this week include the very fast Mercury 3.
Quick answers
What is Mistral Timestral? A family of four AI models from Mistral that forecast time-series data, such as sales, energy use or sensor readings. It was released on October 15, 2026.
Is Timestral open source? The weights are free on Hugging Face under a modified MIT license, and the GitHub code is Apache 2.0. Check the license terms before commercial use.
Does Timestral need a GPU? No. Mistral says all sizes run on a standard CPU, from about 10 ms (5M) to 140 ms (209M) per forecast in its test.
Is Timestral better than TimesFM? Mistral says its 209M model beats TimesFM-3 on public benchmarks with fewer parameters. That is a company claim; watch the GIFT-Eval leaderboard for independent results.
Can I use Timestral in ChatGPT or Claude? No. It is a forecasting model for developers, not a chatbot. You run it in code.
New AI words keep arriving every day. Our agentic AI glossary explains them in plain English.
Sources: Mistral: Timestral · Open Data Science · Hugging Face: Timestral-5M · GitHub: mistralai/timestral
