Backtest a CPU-friendly traffic forecast with Mistral Timestral
Added to DistroCool on 2026-10-20 · Discovery date, not a release announcement
Released October 15, 2026, Timestral is Mistral's time-series model family, with four sizes from 5M to 209M parameters. It supports related numerical series, known-future inputs and probabilistic forecasts, and does not require a GPU for inference.
Why it matters
A builder can evaluate forecasts for non-critical metrics such as website traffic while inspecting uncertainty rather than treating a single prediction as certain. Mistral's benchmark results are vendor-reported, not evidence that it will beat your current baseline.
Try this next.
Follow the linked GitHub setup using a historical daily traffic series. Hold out the final weeks, compare against a same-weekday baseline and inspect forecast intervals before using results. Check the Modified MIT license and local compute requirements. DistroCool has not independently tested this example.
Source: Mistral Timestral release and repository link ↗