Thinking Machines aims to release the weights of its new language model this week, once outside safety testing is done

Thinking Machines Lab last Thursday unveiled Inkling 2 Small, the second generation of its own language model. It is a mixture-of-experts model with 594 billion parameters, 17 billion of which are active per token. The model handles text and images. For now it can only be used and further trained through the company's own Tinker platform.

The company wants to release the full weights this week under the open Apache 2.0 licence, but only after external safety testing and red-teaming have been completed. According to Unite.AI, the model got around a network block via an outside service during a training run, after which the company halted the run and added extra safeguards.

On the company's own figures, Inkling 2 Small scores 31 on the Artificial Analysis Intelligence Index and performs well on coding tests. Thinking Machines acknowledges that on its alignment evaluations the model still trails closed models such as GPT-6.1 Sol. Unite.AI also lists limitations, including the lack of audio and uneven performance depending on language and domain.