Mistral releases Timestral, time-series foundation models with a 209M top model that ranks first on three public benchmarks

Very nice. It used to take me months to build and deploy a forecasting model five years ago, and now I can get apparently quite good results in minutes with an LLM's help and fine-tune on my own data. I want an open weight model with this performance now.

Perhaps I'm misunderstanding, but you should be able to download the models from the Hugging Face collection under mistralai/timestral.

Crazy how they make a 0.03 improvement look like a 5x difference. 0.03 in what units though?

Meh. Time series data is fundamentally different from text, so I see why a time series foundation model sounds appealing to the marketing folks. Real-world usability looks very limited, which is probably why the post uses an imaginary customer instead of a real reference.

Time based data is extremely important, and there are tons of techniques to train super fast, super small models, but they are a pain to tune. I could see a foundation model that is kind of good at a lot of things, then fine-tune it or use it when you want something easier to analyze a bunch of data. It might be very good for feature selection to refine a more specialized model.

I tried Le Chonk for general coding work since they claim it's really good at that, and I was very disappointed. It made all kinds of obvious errors I haven't seen from Opus, Sol, or DeepSeek Flash in over a year.

My field is audio and vision ML, so I don't have domain expertise here, but I don't get how you can lump all time-series together in one bucket. If you're modelling electricity consumption, a simple system that knows the time of day, upcoming holidays, and the weather forecast should beat a model that only sees a window of N previous samples. A 5M model has no idea what a bank holiday is.

You're right, but if the series looks like music, you can apply your music domain knowledge. Part of the magic of large neural nets is that they do recognition and prediction at the same time. Outside info helps even more, but with enough data it should become unnecessary, or so goes the theory.

For some forecasting it doesn't need to be exact. If you're forecasting traffic to scale a load balancer, being a bit off just costs you a little and you adjust after. It basically replaces various traditional curve fitting tools. If you need a really good forecast, you're going to build a domain model.

I interviewed for a company that predicted engine failure on ships from a bunch of sensors on the engine, and they called it just AI. They used LSTMs, and it seemed like a case where good old statistical methods would beat LSTMs. There's no outside information that would help much there.

I think another use is generating the right shape of noisy data. Artificial noise is often added to training data to make ML models more robust, and something like Timestral could simplify generating it.

How does it compare to Nixtla TimeGPT?

Timestral is free and open weight, so there's no need to pay for proprietary TimeGPT.

What's the case for a foundation model when you have 10k different series? Training 10k separate models leaves each with little history, and one generic model has to deal with scale mismatch and feature work across series. A TSFM could save a ton of that work and maybe generalize better.