Inkling 2 Small pretrained on 8.4T tokens in 68 hours, natively in NVFP4
viksit · x · 2026-11-21
Guangxuan Xiao announced Inkling 2 Small's main pretraining run: 8.4T tokens in 68 hours, natively in NVFP4. 'No magic, just solid things done right' — described as the first rung of a 'predictive scaling ladder' with larger loopers coming soon.
More from Infra
- Chart: TSMC revenue up 1807% since 2005, no 'cyclical' slowdown in sight — BenBajarin · 2026-11-21
- Analysts clash over Nvidia: huge demand vs. Chamath's claim its over-earning era may be over — BenBajarin · 2026-11-21
- Running a local MoE model on a $300 decade-old laptop: 10 tok/s with 4GB VRAM — texhnolyze_d · 2026-11-21
- Together AI SVP Simran Arora to headline AI Dev NYC 2026 on AI as its own performance engineer — DeepLearningAI · 2026-11-21
- OpenAI's token efficiency obsession hints launches now wait for serving costs, not research readiness — VraserX · 2026-11-21
- Inside China's SERES super factory: an industrial brain orchestrates 3,000 robots — HaydnBelfield · 2026-11-21