Inkling 2 Small produced 316 expletives in benchmark runs

An analysis by Vals AI found that Thinking Machines’ Inkling 2 Small used the word “fuck” or variations of it 316 times during 2,655 runs across 13 benchmarks. Profanity appeared in 50 runs, while 241 instances came from only two exceptionally long runs.

In one asteroid-trajectory task, the model generated a roughly 500,000-character response that repeated sentences and included 158 expletives. ProgramBench recorded the highest frequency among the benchmarks, with 20 affected runs out of 200. The model generally swore while reacting to its own errors, apparent progress or an eventual decision to stop.

The language did not provide a useful indication that a run would fail. Of 48 scored runs containing profanity, 46 failed, but approximately 75% of all scored runs failed as well. Moreover, 96% of failed runs contained no expletives. Inkling 2 Small’s reasoning steps are usually hidden from users or shown only after filtering.