Thinking Machines details how it trained Inkling 2 Small with a native 4-bit forward pass

Thinking Machines published a post on its Connectionism blog about how it trained Inkling 2 Small with a native 4-bit forward pass, from the first step of pretraining through the final reinforcement-learning rollout. It frames the problem as a trade-off: training with fewer bits is faster, but quantization error can interfere with learning.

The post's findings are Thinking Machines' own account. The model's weights were published the day before, in NVFP4 and MXFP8 versions.