Bittensor's Cascade Beat Salesforce's AI Forecasting Model in Just 19 Days
markjeffrey · x · 2026-11-19
Bittensor's decentralized training subnet Cascade (SN91) has surpassed Salesforce's Moirai 2.0 on GIFT-Eval — a benchmark built by Salesforce AI Research itself covering 144,000 time series and 177 million data points. Cascade hit a CRPS of 0.511 vs Moirai 2.0's 0.5164, just 19 days into its latest training competition, and reached 86.2% parity with Mistral's state-of-the-art Timestral on point-forecast accuracy (MASE). Salesforce trained Moirai 2.0 on 36 million time series (295 billion data points). Mistral's Timestral (Oct 2026, flagship 209M params) currently holds SOTA among zero-shot pretrained forecasting models.
Related event: Bittensor's Cascade MoE Tops GIFT-Eval Benchmark in 19 Days(2 posts)→
More from Models
- micro1 spends $1B a year buying enterprise data; single datasets fetch $2M+ — FinanceYF5 · 2026-11-19
- Clairic open-weights extraction models launch in 27B and 9B sizes — teortaxesTex · 2026-11-19
- New Astra mocked for misreading "all numbers in the paper" as excluding page numbers — analisereal · 2026-11-19
- OpenAI's 6.2-Astra runs 8 hours on AtCoder contest, produces bizarre compressed solution — i_dg23 · 2026-11-19
- Gemini 4 pricing criticized: $7.50-$20 per M output tokens, no cheap or flagship tier — haider1 · 2026-11-19
- Gensyn launches Open8B, the first fully auditable LLM you can verify on a gaming GPU — benfielding · 2026-11-19