Native 4-bit end-to-end training yields just 8% pretraining speedup, sparking debate
andrew_n_carr · x · 2026-11-14
Andrew Carr questions Inkling 2 Small's tech report: the largest model trained end-to-end with native 4-bit forward passes, using unified numerics from pretraining through RL rollout (eliminating quantization calibration). The recipe delivers only an 8% wall-clock pretraining speedup and a 20% RL speedup vs full MXFP8 — surprisingly little for the engineering pain, though RL and inference benefit more.
More from Infra
- SGLang Summit talk details battle-tested NVFP4 4-bit RL across thousands of GPUs — ericzelikman · 2026-11-14
- Anthropic IPO roadshow slide leak sparks Blackwell training speculation — and a block — teortaxesTex · 2026-11-14
- New Yorker interview: AI is a brute-force problem that will 'cook the planet' — kevinnbass · 2026-11-14
- Lenore unveils hybrid Token factory in Zhangjiakou: 15,000P cluster, 2 quadrillion tokens a year — 量子位 · 2026-11-14
- LexiSuite: a Rust-built inference engine beats llama.cpp 1.0.0 at 262k context — W61k3r · 2026-11-14
- Korea Chip Exports Hold Near $2.2B/Day as DRAM Prices Jump 5.5% — Supply Wall or Demand Crack? — tengyanAI · 2026-11-14