Test: Mistral Timestral forecast quality hinges on context length, not model size

A developer benchmarked Mistral's open-weight Timestral time-series models on Paris daily temperatures. With 158 days of context, MAE was 6.489, losing to a repeat-yesterday baseline (5.636); with 1,096 days it fell to 3.326. The 20.8 MB model landed within 0.03 of the 837 MB flagship.

• 158-day context: MAE 6.489, 80% band coverage 43%
• 1,096-day context: MAE 3.326, coverage 64%
• Repeat-yesterday baseline scored MAE 5.636
• 20.8 MB model within 0.03 of the 837 MB flagship