ResearchNovember 16, 2026
-24
🧊

4-Bit Training: Inkling 2 Small Defies Precision Loss

Thinking Machines analyzes 4-bit training and compares BF16, MXFP8 and NVFP4.

#Quantization#Training#NVFP4#LLM#Thinking Machines
4-Bit-Training: Inkling 2 Small trotzt Präzisionsverlust
Share Article
🔥 What happened A team trained Inkling 2 Small, the largest known model with an end-to-end native 4-bit forward pass. It uses NVFP4 for expert layers and MXFP8 for backward passes on Nvidia GB300. 💡 Why it matters NVFP4 stores 3.6x fewer bits than BF16 and runs up to 6x faster. The payoff: 8% faster pretraining wall-clock and 20% faster RL—without wrecking accuracy. ⚡ Our take Training in FP16 now looks like burning compute for sport. But beware: this 4-bit recipe needs specific hardware and custom kernels—not plug-and-play.
The title, summary and analysis of this item were produced automatically by an AI system and have not been editorially reviewed. They may contain errors, bias or omissions — when in doubt, read the linked original source.