Thinking Machines Lab Trains LLM in 4-bit NVFP4
AI-generated summary Thinking Machines Lab published a recipe for end-to-end MoE-LLM training using 4-bit NVFP4 on November 13, 2026. The company claims an 8 percent speedup in pretraining and 20 percent in reinforcement learning compared to MXFP8, with an efficiency loss of approximately 2 percent.