# Timestral Timestral is [[Mistral AI]]'s first family of [[Time Series Foundation Models|time series foundation models]], announced on 15 October 2026. You give it the history of one or more series (sales, sensor readings, energy load...) and it forecasts what comes next. No training on your data first. That's the "zero-shot" part. It's a strange release for Mistral. No chat, no text, no agents. Just numbers in, numbers out. And yet I think it says a lot about where Mistral wants to go: industrial and financial customers who care about running things on their own servers. ## Key facts - **Four sizes:** 5M, 28M, 123M and 209M parameters, same architecture and training recipe. Tiny by [[Large Language Models (LLMs)|LLM]] standards - **Inputs:** multivariate targets, past-only covariates, known-future covariates (a holiday calendar, a planned promotion), and categorical covariates like "machine state" or "day of week", passed as categories instead of fake numbers - **Output:** probabilistic forecasts as 21 quantiles. You don't get one line, you get a range with its uncertainty ([[AI Model Calibration]] matters here) - **Context:** 32,768 time steps - **Hardware:** no GPU needed. On an 8-thread server CPU, forecasting one day from three weeks of 15-minute data takes 17 ms with the 5M model and 113 ms with the 209M one - **Deployment:** on-prem, VPC or air-gapped. Weights on Hugging Face, code on GitHub, and fine-tuning on your own data through Mistral Forge - **License:** "Modified MIT" (more on that below) ## How it's built Mistral's post lists four ideas. Two of them are worth understanding: - **Covariates can't see the future of the target.** Each series gets a role (target, past-only covariate, known-future covariate) and attention only lets information flow from covariates to targets. That blocks a classic leak, where the model "forecasts" a target by peeking at data that wouldn't exist yet - **Mostly synthetic training data.** Mistral generates simple series with trends, seasonality and noise, then wires them together into multivariate series with nonlinear dependencies. Pre-training runs on that with a context of 8,192 steps, then post-training mixes real and synthetic data with a context four times longer The training objective is a pinball loss over the 21 quantiles, a discrete approximation of CRPS (the standard score for probabilistic forecasts). ## What Mistral claims These are vendor numbers, on public benchmarks, with a single checkpoint and no ensembling: - The 209M model ranks first among pre-trained zero-shot models on GIFT-Eval, TIME and fev-bench, on all five metrics Mistral reports (e.g., GIFT-Eval GMean MASE of 0.6575, fev-bench skill score of 50.39%) - The 28M model beats Chronos-2 (Amazon), a model about four times larger, on GIFT-Eval and TIME - The 123M model gets within 0.2% of the 209M score at about three quarters of its latency Public benchmarks are exactly where these models are easiest to overfit, so I'd hold my applause. In the [[Hacker News]] thread, someone from a competing company (The Forecasting Company) claimed a 42% win rate against Timestral on fev-bench. Also a vendor number. Test on your own data. ## The license catch "Modified MIT" sounds like MIT. It isn't. The license removes your rights once you go above a revenue threshold of $20M per month. Several commenters called the label misleading, and I agree: if the license has a revenue cap, say so in the name. For most of us it changes nothing, but if you work for a large company, read the license before you build on it. Same lesson as with any [[AI Open Weight Models|open-weight model]]: open weights don't mean open source. ## What the Hacker News discussion added The thread (600+ points) was more about time series foundation models in general than about Timestral. I captured that part in [[Time Series Foundation Models]]: classical baselines, the scale argument, covariates, and how to evaluate a forecaster without fooling yourself. Two points were specific to Timestral: - **People actually ran it on a CPU.** One commenter confirmed it works without a GPU, which matches Mistral's claim - **It's a black box.** A signal processing engineer asked whether you can extract the underlying dynamics from it. You can't. You get forecasts and quantiles, not a model of the system ## My take Mistral's own example is the interesting one: Timestral flags an anomaly in a sensor stream, an agent compares it with past failures, and sub-agents dig through maintenance logs and manuals to find the root cause. That's the pattern I keep seeing (see [[Cloudflare Clairic]] and [[TabFM]]): small, specialized models that return typed values, with an LLM-based [[AI Agents|agent]] only where you need language and judgment. The LLM isn't the forecaster. It's the colleague who reads the forecast. For a European industrial company that wants to keep its data at home ([[Sovereign AI]]), a forecaster that runs on a CPU in an air-gapped plant is a compelling offer. Would I pick it over a tuned classical model for one important series? No. Would I try it as the default for a thousand series nobody has time to model? Absolutely. ## References - Mistral announcement, "Timestral: Mistral's time-series foundation models" (15 October 2026): https://mistral.ai/news/timestral/ - Hacker News discussion: https://news.ycombinator.com/item?id=50150434 ## Related - [[Time Series Foundation Models]] - [[Mistral AI]] - [[AI Foundation Models]] - [[TabFM]] - [[Cloudflare Clairic]] - [[AI Model Calibration]] - [[AI Open Weight Models]] - [[MIT License]] - [[Sovereign AI]] - [[Machine Learning (ML)]] - [[AI Agents]] - [[Hacker News]]