# Mistral's Timestral forecasting models run from 5M to 209M parameters on CPU

> Mistral's first time-series foundation models claim top zero-shot scores on GIFT-Eval, TIME and fev-bench, and take 10 to 140 ms for a one-day forecast on an 8-thread CPU.

- Published: 2026-10-19
- By: Samir Sengupta (https://www.samcodeman.com/about)
- Canonical: https://www.samcodeman.com/writing/mistral-timestral-forecasting-models-5m-209m-cpu
- How this was made: Drafted with Claude, Anthropic's model from 1 cited source, then checked against it in a second editing pass. Editorial policy: https://www.samcodeman.com/editorial-policy

## The short version

- Mistral reports that Timestral ranks first among pre-trained zero-shot models on all five metrics across GIFT-Eval, TIME and fev-bench, using a single checkpoint without ensembling.
- The 28M model is five times faster than the 209M model. Mistral says it beats Chronos-2, about four times larger, on GIFT-Eval and TIME. The 209M model beats TimesFM-3 and Falcon-2.0 with fewer parameters.
- Covariates are assigned explicit roles (target, past-only or known-future), so calendars and weather forecasts can feed a forecast without leaking a target's own future.
- The licence is listed only as Modified MIT, and the release post does not spell out the modifications, so check the licence text before production use.

Mistral released Timestral on October 15, 2026. It is a family of four time-series foundation models at 5M, 28M, 123M and 209M parameters, with a 32,768-step context, probabilistic output over 21 quantiles, and a Modified MIT licence. Mistral says the 209M model ranks first on GIFT-Eval, TIME and fev-bench with two-thirds the parameters of the previous state of the art. It also says the 28M model beats Chronos-2, about four times its size, on GIFT-Eval and TIME. Inference does not need a GPU: on an 8-thread server CPU, a one-day-ahead forecast from three weeks of 15-minute history takes 10 ms with the 5M model and 140 ms with the 209M model.

The models take multivariate targets plus past-only and known-future covariates, including categorical ones such as day of week or machine state. Mistral lists on-prem, VPC and air-gapped deployment, has published a GitHub repository for running the models, and offers fine-tuning through Mistral Forge. Mistral says a technical report with full results will follow; it has not yet been published. Every benchmark claim below comes from Mistral's own post and has not been independently reproduced.

## What the four Timestral sizes trade off

All four sizes share one architecture and training recipe. Larger models forecast better, and smaller ones are faster and cheaper. Mistral's latency figure covers a one-day-ahead forecast from three weeks of 15-minute history. That works out to 2,016 input steps and 96 output steps, a small slice of the 32,768-step context. On that task the 5M model takes 10 ms and the 209M model takes 140 ms on a standard server CPU with 8 threads.

- 28M: five times faster than 209M. Mistral says it beats Chronos-2 on GIFT-Eval and TIME.
- 123M: within 0.2% of the 209M score at three quarters of its latency.
- 209M: the top-ranked checkpoint, with about two-thirds the parameters of the previous state of the art.
- 5M: 10 ms per forecast on the stated task, the fastest and cheapest of the four.

For a plant with thousands of sensors, the 123M model is the obvious tradeoff to test first. It gives up 0.2% of the 209M score for a 25% latency cut. The post does not report memory footprint, batch throughput, or latency at the full 32,768-step context. Teams sizing CPU fleets will need to measure those themselves.

## How Timestral compares with Chronos-2 and TimesFM-3

Mistral reports results on three leaderboards:

- GIFT-Eval: GMean MASE and GMean WQL, top ten shown out of 83 comparable pre-trained zero-shot entries.
- TIME: GMean MASE and GMean CRPS, top ten shown out of 35 leaderboard entries.
- fev-bench: SQL skill score, top ten shown out of 37 leaderboard entries.

Across those five metrics, Mistral says Timestral is first among pre-trained zero-shot models, from a single checkpoint without ensembling. The specific head-to-head claims are that the 28M model beats Chronos-2 on GIFT-Eval and TIME, and that the 209M model beats TimesFM-3 and Falcon-2.0 with fewer parameters. The 28M claim covers GIFT-Eval and TIME, not fev-bench.

The post's text gives no exact scores, only charts. It says results have been submitted to the GIFT-Eval and TIME leaderboards, and fev-bench is not mentioned in that sentence. Until those entries appear and the technical report lands, the margins over Chronos-2 and TimesFM-3 cannot be read precisely. Public benchmark wins also do not settle production fit, a point Mistral makes itself when it argues for in-domain fine-tuning.

## Weights, licence and how to run Timestral

The spec table lists the licence as Modified MIT. The post does not say what the modifications are, which is the first thing to check before shipping the model inside a commercial product. For running the models, Mistral says it has published a GitHub repository. The announcement text does not include its URL or usage code, so no code example is shown here.

On weights, the post itself is not explicit. When one commenter on Hacker News asked for an open-weight model with this performance, another replied that the models can be downloaded from a Hugging Face collection at huggingface.co/collections/mistralai/timestral. That is a commenter's pointer, not a statement in Mistral's post. Confirm the files and the licence text there directly.

## Known-future covariates are a first-class input

One Hacker News commenter who works in audio and vision ML questioned lumping all time series into one bucket. They argued that knowing the time of day and upcoming holidays should be the basis of an electricity forecast, not an optional improvement. Timestral's architecture handles part of this by giving each input series a role: target variate, past-only covariate, or known-future covariate. The post describes the rule as:

> Known-future covariates are visible across the whole window while targets stay causal, and information flows only from covariates to targets.
>
> - Mistral, release post

Categorical covariates are taken as categories rather than numbers, so machine states such as running, idle or under maintenance are not given a false ordering. Multivariate targets are modeled jointly. Mistral's example is a pump whose vibration, temperature and pressure inform one another. Another commenter described a ship-engine failure model built only on engine sensors, with no external inputs, and suspected classical statistical methods would beat the LSTMs that company used. The post does not show ablations quantifying how much covariates or joint modeling improve accuracy.

## Decoupled horizon forecasting keeps history as context

Mistral contrasts its training approach with causal decoders such as Toto 2.0 and TimesFM-3, which use contiguous patch masking. In that approach, values to be forecast are masked out, so they look like missing data and are lost as context for later forecasts. Decoupled Horizon Forecasting (DHF) keeps the full history in the sequence and adds learned forecast tokens that read the history before them. Every observed value serves as both target and context, and true gaps stay flagged as missing. Mistral also says this lets training horizon lengths and missing-data augmentation be tuned independently, which is relevant for production data with sensor dropouts.

Training is synthetic-first, and the majority of the training data was synthetic. Mistral generated univariate processes with trends, seasonality and varied noise, then built multivariate series by adding nonlinear dependencies between them. It pre-trained on purely synthetic data with context 8,192, then post-trained on a mix of real and synthetic data with a context four times longer, which matches the 32,768-step limit. The objective is a pinball loss over 21 quantiles, a discrete approximation of CRPS. The composition of the real-data mix is not disclosed, which matters for anyone worried that their benchmark or backtest data overlaps with training data.

## Fine-tuning through Mistral Forge for drifting data

One Hacker News commenter called real-world usability very limited and noted that the post features an imaginary customer rather than a reference. That criticism holds: the industrial plant in the post is hypothetical. A commenter identifying as a cofounder of The Forecasting Company disagreed, saying time-series foundation models are in production at quant funds and top retail companies. That is a claim about the category, not about Mistral's models. Other commenters placed such models where a rough forecast is acceptable, such as scaling a load balancer, or suggested using them for feature selection ahead of a specialized model.

Mistral's own position is that public benchmarks measure general performance, and that lasting value comes from adapting to in-domain data as distributions shift. Mistral Forge, the same pipeline Mistral uses for its language models, supports:

- zero-shot inference out of the box
- architectural extensions such as domain-specific covariates
- fine-tuning on proprietary data while preserving general performance
- control over the training mix, including replaying pre-training data and generating synthetic samples for the target domain

Mistral says control over the data ensures nothing it backtests was seen during pre-training. It also says the full pipeline can run on client hardware, including air-gapped environments. The post gives no Forge pricing, no fine-tuning compute requirements, and no before-and-after accuracy numbers for a fine-tuned model.

## Questions this raises

### Does Timestral need a GPU to run?

No. Mistral says a one-day-ahead forecast from 2,016 input steps takes 10 ms with the 5M model and 140 ms with the 209M model on an 8-thread server CPU. The post does not report memory footprint, batch throughput, or latency at the full 32,768-step context.

### Have Timestral's benchmark results been independently verified?

No. Every benchmark claim comes from Mistral's own post, which shows charts but no exact scores. Mistral says results were submitted to the GIFT-Eval and TIME leaderboards and that a technical report will follow.

### What license does Timestral ship under?

The spec table lists a Modified MIT licence, but the post does not say what the modifications are. Check the licence text before shipping the models in a commercial product.

## Key facts

- Mistral released Timestral on October 15, 2026 in four sizes: 5M, 28M, 123M and 209M parameters.
- A one-day-ahead forecast from three weeks of 15-minute history takes 10 ms (5M) to 140 ms (209M) on an 8-thread server CPU.
- The 123M model is within 0.2% of the 209M score at three quarters of its latency.
- All sizes support a 32,768-step context and output 21 quantiles under a Modified MIT licence.
- Mistral says the 28M model beats Chronos-2 on GIFT-Eval and TIME, but the technical report with exact scores is not yet published.

## Sources

1. [Timestral: Mistral release time-series foundation models](https://mistral.ai/news/timestral/) - mistral.ai

Discussion read (practitioners' reactions, not a source): https://news.ycombinator.com/item?id=50150434