Otto365 AI Brief #018
Mistral: forecasts as a range, not a single guess
What happened
Mistral's new forecasting model gives a range, not a single guess. Mistral AI announced Timestral, its family of time-series forecasting models, on the 15th of October 2026. They forecast sensor readings or finance flows and run on a CPU.1
Three things to know
- Four sizes, 5M to 209M First, four sizes, 5 million to 209 million parameters, and Mistral AI says the largest ranks first on GIFT-Eval, TIME and fev-bench. Day-ahead forecasts take 10 milliseconds on the smallest, 140 on the largest.1
- Range, not one number Second, each forecast is a range of 21 quantiles, not one number, and related series such as vibration and temperature are modelled together. Known events such as planned shutdowns or weather forecasts feed in too.1
- GitHub, modified MIT Third, a GitHub repository lets you run it on your data under a modified MIT licence. Fine-tuning goes through Mistral Forge, its own training pipeline, which can run on your hardware, even air-gapped.1
Why it matters to you
Otto365's take (opinion)
So, my take. Forecasting with a range is the useful idea for a small firm planning stock or staff, not a plug-in app yet. Mistral AI's example pairs it with agents that turn an alert into a diagnosis.1
Do this week
Try this: mark known events, like holidays, in a year of sales.
Sources
- Mistral AI: Timestral: Mistral's time-series foundation models | Mistral (opens in a new tab) Primary source, accessed 16 Oct 2026
The quotes we checked (27)
October 15, 2026 By Mistral
Mistral AI · source 1
Today, we’re releasing Timestral, our first family of time-series foundation models.
Mistral AI · source 1
Our largest model, at just 209M parameters, ranks first on all three of the field's main public benchmarks, GIFT-Eval, TIME, and fev-bench
Mistral AI · source 1
Available in four sizes (5M to 209M parameters), it balances accuracy and efficiency, running efficiently on CPUs
Mistral AI · source 1
Timestral does not need a GPU at inference.
Mistral AI · source 1
On a standard server CPU with 8 threads, forecasting one day ahead from three weeks of 15-minute history takes 10 ms with the 5M model and 140 ms with the 209M model.
Mistral AI · source 1
The larger models improve forecast performance, while the smaller ones are faster and cheaper.
Mistral AI · source 1
The objective is a pinball loss over 21 quantiles, a discrete approximation of CRPS.
Mistral AI · source 1
Timestral predicts a full set of quantiles, so teams can see how wide the uncertainty is, set alert thresholds that reflect it, and plan maintenance around risk rather than a single number.
Mistral AI · source 1
Timestral models related series jointly, so it can use those dependencies instead of forecasting each sensor in isolation.
Mistral AI · source 1
A pump's vibration, temperature and pressure don't move independently.
Mistral AI · source 1
Some drivers of sensor behaviour are known in advance, such as production schedules, planned shutdowns, weather forecasts.
Mistral AI · source 1
Timestral can therefore use a holiday calendar or a weather forecast without leaking a target's own future into its forecast.
Mistral AI · source 1
We have published a GitHub repository so you can run Timestral on your own data.
Mistral AI · source 1
Deployment On-prem, VPC or air-gapped Licence Modified MIT
Mistral AI · source 1
Timestral is available through Mistral Forge, the same training pipeline we use for our language models.
Mistral AI · source 1
Fine-tuning on proprietary data, specialising to a domain while preserving general performance.
Mistral AI · source 1
The full pipeline can be deployed on our clients’ hardware, including air-gapped environments.
Mistral AI · source 1
Timestral is a numerical model that can run inside existing monitoring pipelines, or be paired with our language models in multi-modal agentic workflows.
Mistral AI · source 1
The plant manager gets a diagnosis and a plan backed by the sensor data, rather than just an alert.
Mistral AI · source 1
When it flags an anomaly on a pump, an agent embeds the signals and compares against embeddings of past sensor data to find similar historical failures.
Mistral AI · source 1
Time series is one of the most common data types in enterprise, from sensor telemetry to observability metrics.
Mistral AI · source 1
Finance teams forecast flows, volumes, and volatility to inform positions, hedging, and treasury across thousands of accounts.
Mistral AI · source 1
Among pre-trained, zero-shot models, it ranks first across all five metrics. These results come from a single checkpoint, without ensembling.
Mistral AI · source 1
The agent then starts sub-agents to run a root cause analysis, drawing on context from that plant, such as maintenance records, shift logs and operator notes
Mistral AI · source 1
Timestral running continuously on the plant's sensor streams, detecting anomalies as they occur.
Mistral AI · source 1
Timestral delivers probabilistic, multivariate forecasts and supports past-only and known-future covariates.
Mistral AI · source 1
Transcript
Full transcript
Mistral's new forecasting model gives a range, not a single guess.
Play from 0:04 Mistral AI announced Timestral, its family of time-series forecasting models, on the 15th of October 2026. They forecast sensor readings or finance flows and run on a CPU.
Play from 0:18 First, four sizes, 5 million to 209 million parameters, and Mistral AI says the largest ranks first on GIFT-Eval, TIME and fev-bench. Day-ahead forecasts take 10 milliseconds on the smallest, 140 on the largest.
Play from 0:35 Second, each forecast is a range of 21 quantiles, not one number, and related series such as vibration and temperature are modelled together. Known events such as planned shutdowns or weather forecasts feed in too.
Play from 0:49 Third, a GitHub repository lets you run it on your data under a modified MIT licence. Fine-tuning goes through Mistral Forge, its own training pipeline, which can run on your hardware, even air-gapped.
Play from 1:03 So, my take. Forecasting with a range is the useful idea for a small firm planning stock or staff, not a plug-in app yet. Mistral AI's example pairs it with agents that turn an alert into a diagnosis.
Play from 1:17 Try this: mark known events, like holidays, in a year of sales.
I'm Otto365.
Follow the Brief.
On screen
- Play from 0:00
- Otto365 AI Brief #018, Mistral, 15 Oct 2026
- Forecasts as a range, not a single guess
- Past readings
- Known events
- Timestral
- Range of outcomes
- Play from 0:04
- Timestral: forecasting
- Announced 15 Oct 2026
- Timestral
- Four sizes, 5M to 209M
- Runs on CPU, no GPU
- 21-quantile forecasts
- Modified MIT licence
- Source: Mistral AI, 15 Oct 2026
- Play from 0:18
- Four sizes, 5M to 209M
- 5M model: 10 ms a forecast; Faster, cheaper
- 209M model: 140 ms a forecast, Ranks first
- First on all three benchmarks
- Pre-trained, zero-shot models
- One checkpoint, no ensembling
- Play from 0:35
- Range, not one number
- Related sensors
- Modelled jointly
- 21 quantiles
- Alert thresholds
- Planned shutdowns
- Weather forecasts
- Holiday calendar
- Play from 0:49
- GitHub, modified MIT
- Your own data
- GitHub code
- Mistral Forge
- Your hardware
- Fine-tune on your own data
- On-prem, VPC or air-gapped
- Play from 1:03
- Range beats one guess
- One number: A single guess
- A range: 21 quantiles, Plan around risk
- Sensor streams
- Anomaly flagged
- Root cause check
- Diagnosis, plan
- Play from 1:17
- Mark known events in sales
- Voiced by AI. Checked by humans. Apps 365 sells AI and Microsoft 365 services. Not endorsed by companies named.
Voiced by AI. Checked by humans. Apps 365 sells AI and Microsoft 365 services. Not endorsed by companies named. Report an error.