Timestral Mistral Time Series LLM

Author: Data Science in your pocket · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Timestral Mistral Time Series LLM · Channel: Data Science in your pocket · 07:16 · Watch on YouTube

Topics: AI agents, AI evaluation and benchmarks, Model inference and optimization, AI models

Summary

The video introduces Timestral, a family of four time series foundation models ranging from 5 million to 209 million parameters, and reports that it ranks first across five metrics on three public forecasting benchmarks in zero-shot evaluations. It emphasizes the models’ efficiency, including CPU inference, and describes features such as multivariate and probabilistic forecasting, covariates, and a synthetic-first training process. The video also outlines possible use in agentic systems and customization through Mistral Forge. It cautions that the release is still on day one, with details such as the full technical report, API pricing, and deployment case studies not yet available.

Key points

Timestral consists of four models with 5M, 28M, 123M, and 209M parameters, sharing the same architecture and training recipe.
The video reports first-place zero-shot results across five metrics on the Gift, Time, and Feb benchmarks.
The benchmark results are described as coming from a single checkpoint and a single run, without ensembling.
The 28M model is reported to outperform Chronos 2, while the 209M model is said to match prior state of the art with roughly two-thirds the parameters.
Timestral can run forecasting on a standard eight-thread CPU, with reported inference times of 10 ms for the 5M model and 140 ms for the 209M model.
Its capabilities include multivariate forecasting, 21 quantiles, categorical covariates, and known future covariates.
The architecture uses learned prediction tokens, covariate-aware attention, and a quantile head trained with pinball loss.
Training starts with synthetic data and then uses a mix of real and synthetic data with a longer context.
Mistral presents Timestral as a potential component in agentic systems and offers customization through Mistral Forge.
The video notes that the modified MIT license is unexplained and that pricing, memory figures, the full technical report, and customer case studies are not provided.

Chapters

Start	Chapter	Summary
00:00	Introducing Timestral and time series foundation models	The video explains the role of time series forecasting, introduces Timestral’s four model sizes, and presents its benchmark claims.
02:03	Benchmark results and efficiency	It reviews reported benchmark scores, compares Timestral’s accuracy against model size, and describes CPU inference speeds.
04:07	Forecasting features and model design	The video covers multivariate and probabilistic forecasting, covariates, the architecture’s attention and prediction approach, and its training recipe.
05:42	Agents, customization, and caveats	It discusses agentic-system use and Mistral Forge, then lists missing release details and limits on the benchmark claims.
06:46	Release-day takeaway	The video concludes that Timestral is an impressive early release and recommends it as worth exploring for forecasting work.

Quotes

Not billion, million. — 00:00

One model, one run, top of the table. — 02:03

Timestral doesn’t need a GPU at all. — 03:35

It’s synthetic first. — 05:42

This is day one, impressive, but day one. — 07:16

Concepts and entities

Timestral, Mistral, Mistral Forge, ARIMA, PatchTST, TimesFM 3.0, GIFT-Eval, Time benchmark, Fев Bench, Hugging Face

Full transcript (original language)

00:00 209 million parameters. Not billion, million. And this tiny model just took the number one spot on all three of the world’s main time series forecasting benchmarks. This is Time Mistral’s first ever family of time series foundation models, released October 15th, 2026. And it might be the most efficient foundation model release of the year. Let’s break it down. Time series is quietly the most common

00:30 data type in the enterprise. Sensor telemetry, server metrics, electricity prices, sales figures, it’s all time series. And forecasts built on it drive real decisions, capacity planning, energy trading, supply chains, predictive maintenance. For decades, classical models like ARIMA learn from one series at a time. Then deep models like PatchTST learn richer patterns, but still had to be trained from scratch on every new data set. No transfer.

01:01 Time series foundation models changed that. Pre-train once on a huge, diverse corpus, then forecast data they’ve never seen in zero-shot. Time-ASTRAL isn’t one model, it’s four. 5 million, 28 million, 123 million, and 209 million parameters. All four share the exact same architecture and training recipe. You just pick your point on the accuracy versus speed curve. And Mistral claims every single size sits on the Pareto front at each

01:31 parameter count. No public model forecasts better. The smallest one, 5 million parameters, is genuinely tiny, small enough to live inside a sensor gateway. Now, the claims. Mistral evaluated Time-ASTRAL on the three main public forecasting benchmarks, Gift N Bell, Time, and Feb Bench. Among pre-trained, zero-shot models, Time-ASTRAL ranks first across all five metrics. And here’s the part one like these results come from a single checkpoint with no ensembling tricks.

02:03 No averaging 10 runs together. One model, one run, top of the table. Let’s read the scoreboard. On Gift Val point accuracy, geometric mean maze, Timestral scores 0.657, edging out T0 gamma and Google’s Times FM3. On the probabilistic metric, WQL 0.446, again first. On the time benchmark, 0.626 maze and

02:35 0.522 CRPS, both number one. And on Feb bench is still score, where higher is better, 50.4 ahead of Times FM3 and Amazon’s Chronos 2. Five metrics, five wins against Google, Amazon, Huawei, LG, and IBM’s entries. But raw leaderboard wins aren’t the real story, efficiency is. This chart plots forecast error against model size, log scale, and the orange line is the Timestral family.

03:05 It sits below the entire previous Pareto front. Translate that, the 28 million parameter Timestral beats Chronos 2, a model about four times larger. The 209M beats Times FM3 and Falcon 2.0 with fewer parameters. Mistral says the largest model matches the previous state of the art with roughly 2/3 the parameters, though they don’t name exactly which model that’s measured against.

03:35 And here’s the number that matters for deployment, Timestral doesn’t need a GPU at all. On a standard server CPU with eight threads, forecasting one day ahead from three weeks of 15-minute history takes 10 milliseconds with the 5M model and 140 milliseconds with the 209M. The 28M is five times faster than the largest. And the 123M comes within 0.2% of the top score at 3/4 of the latency. So, what can it actually do? Four things worth knowing.

04:07 One, multivariate forecasting. It models related series jointly like pump vibration, temperature, and pressure together. Two, probabilistic forecasts. Not just one line, but 21 quantiles. A full uncertainty band for risk-aware decisions. Three, categorical covariates out of the box. Day of week, machine state, running or idle, no manual encoding. And four, known future covariates. Feed it the production schedule or the weather forecast, and a planned shutdown stops looking like an anomaly.

04:40 The blog’s live demo shows exactly this on German electricity prices with load and solar and wind as covariates. Under the hood, three ideas matter. First, decoupled horizon forecasting. Models like TimesFM3 mask the values they’re supposed to predict, which makes them look like missing data. Timestral instead keeps the full history in context and uses learned prediction tokens for the future patches. Every observed value serves as both target and context.

05:11 Second, covariate-aware attention. Every input series gets a rolling target, past only covariate, or known future covariate, and information flows only from covariates to targets. So, it can use a holiday calendar without leaking the future. Third, a quantile head trained with pinball loss. The training recipe is the other surprise. It’s synthetic first. The majority of Timestral’s training data wasn’t real-world data at all. It was generated. Univariate processes with

05:42 trends, seasonality, and diverse noise. Then a multivariate synthesizer added non-linear dependencies between series. Training happens in two stages, pre-training on purely synthetic data with an 8,000 step context, then post-training on a mix of real and synthetic data with four times longer context, 32,768 time steps in total. Mistral’s bigger play, Timestral as a component in agentic systems. Their example, Timestral monitors plant

06:12 sensor streams and flags anomalies. A language model agent compares the signal against past failures, sub-agents dig through maintenance logs and manuals, and the plant manager gets a diagnosis and a plan, not just an alert. And when benchmarks aren’t enough, there’s Mistral Forge, the same training pipeline behind their language models, zero-shot out of the box, architectural extensions, fine-tuning on your proprietary data, full control over the training mix, and the whole pipeline can run on your own hardware, even air-gapped.

06:46 Now the fine print, because there’s always fine print. The license is modified MIT, and the blog never says what was modified. There’s no API pricing, no endpoint name, no memory footprint numbers, and the full technical report will follow. It’s not out yet. The number one claims are scoped to pre-trained zero-shot models, and the charts show top 10s, not full leaderboards. No named customers, no deployment case studies.

07:16 This is day one, impressive, but day one. So a 209 million parameter model, number one on every major benchmark, running on a CPU in milliseconds, with weights on Hugging Face and a GitHub repo today. For anyone doing forecasting, this is worth an afternoon. That’s Timestral. If this breakdown helped, subscribe to Data Science in Your Pocket, links below. See you in the next one.