Mistral's Timestral tops time-series benchmarks with 209M params, built mostly by agents
hkproj · x · 2026-10-24
A small applied-science team at Mistral built Timestral as a side project: #1 among pretrained zero-shot models on GIFT-Eval, TIME and fev-bench. The 209M model beats previous SOTA with 2/3 the parameters, runs on CPU in milliseconds, and weights are open.
Agents did most of the work, with Mistral Large 4 as a major contributor — building synthetic data pipelines, suggesting architecture changes, and hill-climbing experiments overnight on Forge, Mistral's training platform. The company also notes ML4 was trained on just 3,800 GPUs, far fewer than other frontier labs.
More from AGI Musings
- Runway CEO: generative AI quality, resolution and control double roughly every year — c_valenzuelab · 2026-10-24
- dhh: China's Programmers' Day shows why its engineering culture wins at AI — prasanna_says · 2026-10-24
- Founder: The Best AI Use Is Augmenting Your Best People, Not Full Agent Autonomy — jacob_posel · 2026-10-24
- Cloudflare CEO: 'Every single week it's going to feel like the sky is falling' as AI uncovers critical vulnerabilities — Polymarket · 2026-10-24
- Svante Pääbo's team uses AlphaGenome as a time machine to reconstruct human gene regulatory evolution — gleech · 2026-10-24
- KAN author Ziming Liu: the "language of AI science" hasn't been invented, so LLMs can't learn it — ZimingLiu11 · 2026-10-24