Inkling 2 Small pretrained on 8.4T tokens in 68 hours, natively in NVFP4
teortaxesTex · x · 2026-11-07
- Guangxuan Xiao's team announced the main pretraining phase of Inkling 2 Small is done: 8.4T tokens in just 68 hours, trained natively in NVFP4.
- The team says there's no magic, just solid things done right, and calls it one rung of their "predictive scaling ladder" with more to come.
- Amplifiers speculate the run is highly efficient and wonder about MFU and the number of Blackwell GPUs used (not disclosed); some call it possibly the first seriously published natively FP4-trained model.
More from Infra
- Crusoe Sells Troubled Wyoming Data Center Project to Google — TechCrunch AI · 2026-11-07
- SemiAnalysis says tighter GPU financing supports rents, not a glut — analysts push back — bookwormengr · 2026-11-07
- CorebrumOS, a Physical AI OS, now installs on any Ubuntu device including MacBook Pro — chrismatthieu · 2026-11-07
- Beam report also covers optimization, training dynamics and hyperparameter transfer — brandondamos · 2026-11-07
- Copper prices fall as markets fear an AI backlash, reports Financial Times — nordicinst · 2026-11-07
- macOS CI at $0.16/min pays for a Mac Mini in 10 days — why CI prices should drop 50%+ — gajesh · 2026-11-07