Thinking Machines Lab published a post on its assessment of the Inkling model. The company states that releasing weights indiscriminately is unsafe and that keeping capable models inside a few labs is also unsafe. It presents staged access as the visible part of a middle path, with testing and expanding safety work as defenses mature. The post is hosted on the company's site. Replies from AI researchers and open-source figures note the post's discussion of marginal risk and pretraining data filtering.