Thinking Machines Lab describes end-to-end 4-bit training of Inkling 2 Small

Thinking Machines Lab trained Inkling 2 Small end-to-end with a native 4-bit forward pass, using NVFP4 in MoE expert layers for the forward pass and MXFP8 with stochastic rounding for backward passes. The lab says the recipe delivered an 8% wall-clock speedup in pretraining and 20% in RL over full MXFP8 training, at a cost of about 2% compute efficiency in pretraining, and that Inkling 2 Small was pretrained in 68 hours on GB300 NVL72 racks.