HeyHomees Lab Notes · No. 5 · Mistral · Published October 20, 2026
By HeyHomees Newsroom ·
In 30 seconds
- Four model sizes (5M-209M) forecast multivariate series with 21 quantiles per step.
- At launch Timestral-209M ranked #1 on GIFT-Eval, TIME, and fev-bench zero-shot benchmarks.
- The code is Apache-2.0; model weights are under a Modified MIT license and can run on CPU
What is it?
Timestral is a family of time-series foundation models for multivariate probabilistic forecasting. It supports past-only, known-future, and categorical covariates, and outputs 21 non-crossing quantiles per time step.
How does it work?
The architecture is a causal patch-based transformer trained with Decoupled Horizon Forecasting, using separate prediction tokens that read a causally encoded context. Forecasts are generated block-by-block (32 patches = 1024 steps) with each block's median fed back as context.
How well does it do?
At launch Timestral-209M ranked first among pretrained and zero-shot models on three public forecasting benchmarks: GIFT-Eval (#1), TIME (#1 across 98 tasks), and fev-bench (#1 across 100 tasks). All four sizes sit on the Pareto front of accuracy versus model size.
What does it need to run?
The repository requires only PyTorch, NumPy, safetensors, and huggingface_hub. Inference can run on CPU or GPU, and the context is capped at 32,768 steps.
By the numbers
- Model parameter counts: 5M, 28M, 123M, 209M
- Maximum context length: 32768 steps
- Number of quantiles per time step: 21
- Decode block size: 1024 steps
The honest read
The reported #1 rankings come from Mistral AI’s own zero-shot evaluation on three public benchmarks; no external comparison or statistical significance is provided. The figures show model size and weight size, but they do not convey inference speed, memory usage beyond the context limit, or performance on domain-specific data. The code runs on CPU, yet the benchmarks likely used GPU-accelerated inference, which is not disclosed. The model weights are released under a Modified MIT license separate from the Apache-2.0 code, meaning users must respect two different licensing terms. The Pareto-front claim reflects a trade-off between accuracy and size, but absolute accuracy numbers are not given.
What we don't know yet
- Exact inference latency and memory consumption on CPU vs GPU
- Training data sources and preprocessing details
- Whether the zero-shot benchmark scores are reproducible by third parties
Sources
- README, GitHub (primary)
Researched and drafted with AI from the sources above; every figure is checked against them by code before it is published, and a line that fails is dropped.