Mistral shipped a time-series foundation model today. I measured its latency — and its confidence.

Description and use

Latency scales linearly with thread count; 5M model meets 10 ms claim on 8 threads, but requires ~50 ms on 2 threads. Point forecasts outperform seasonal naive by 37% MAE. Probabilistic intervals are severely under‑covered (17% for 80% band), making them unreliable for alerting without local calibration.

Problem

Assessing the real-world performance of Mistral’s new Timestral time-series foundation models on CPU latency, point forecast accuracy, and probabilistic calibration.

Outcome

Workflow

1. Download Timestral‑5M and Timestral‑28M weights from Hugging Face.
2. Verify parameter counts from weight files.
3. Run inference on a 2‑vCPU VM with torch pinned to 2 threads, measuring median latency over seven runs.
4. Forecast 96 steps ahead from 2,016‑step context (three weeks at 15‑min intervals).
5. Compare latency to Mistral’s 8‑thread claim.
6. Evaluate point forecast MAE on monthly airline passengers (1949‑1960) against seasonal naive and drift baselines.
7. Compute 80% prediction interval coverage and bias over 24‑month hold‑out.
8. Analyze quantile monotonicity and interval width growth.
9. Summarize findings and recommend capacity planning and calibration checks.

Evidence

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.

Source

model Mistral claim (8 threads) mine, median (2 threads) mine, min Timestral-5M 10 ms 52.4 ms 35.9 ms

Source

MAE (passengers, thousands) Timestral-5M median 44.93 seasonal naive 71.25 drift 91.62

Source

80% interval coverage, 24 months 4 of 24 = 16.7% first 12 months 25.0% last 12 months 8.3% median bias over 24 months -44.93

Source

Record date: 2026-10-20T00:41:10.854Z