Thinking Machines Releases Inkling-Small, Matching Inkling at a Quarter of the Size

US-based AI company Thinking Machines Lab released its open-source model Inkling-Small on July 30, which performs nearly as well as its flagship Inkling but at a quarter of the size. Using a mixture-of-experts architecture with 276B total parameters and 12B active ones, it was trained on NVIDIA GB300 NVL72 systems and supports a 1M token context window along with native audio and image reasoning. In tests, Inkling-Small showed higher efficiency in agent tool use, reasoning, and instruction following, scoring 31.6% on Humanity's Last Exam, beating Inkling's 29.7%. The company has released full weights and integrated it into Tinker for fine-tuning and Playground for text, image, and audio chat experiences.