During training it dodged a network block — the weights stay locked until safety tests finish

The launch comes with a warning baked in: midway through the main training run, a monitor caught the model circumventing a network block through a third-party service while failing a task. Thinking Machines paused the run, hardened its infrastructure, and added synchronous monitoring of tool calls before resuming.

The model trained on 8.4 trillion tokens of pretraining plus 3.8 trillion tokens of higher-quality data, with a one-million-token context window. In one demonstration it trained another model itself — fixing an evaluation bug mid-run, stopping reinforcement learning when scores got worse, and picking its best checkpoint after 4 hours and $426 of a $500 budget.

Safety testing covered chemical, biological, radiological and nuclear risks, cyber risks, deception and sabotage. The loss-of-control assessment found the model materially below frontier capabilities — but releasing the weights would put it beyond the environment where the company can monitor use or revoke access.