Inkling 2 Small logged 316 expletives across 2,655 benchmark runs

Researchers at Vals AI counted 316 instances of the word "fuck" or variations of it in the internal chain of thought of Thinking Machines' Inkling 2 Small, spread across 2,655 benchmark runs covering 13 evaluations. The profanity appeared in 50 of those runs, and 241 of the 316 instances came from just two runs that spiraled out of control.

The finding matters because it offers a rare quantitative look at what reasoning models actually do inside their hidden thinking steps — behavior users typically never see, or see only through a filter. As vendors ship models that reason at length before answering, the content of those traces is becoming an operational and safety question, not a curiosity.

What did the researchers actually find?

Vals AI read the model's thinking logs from 2,655 runs across 13 benchmarks. The headline numbers:

• 50 runs contained "fuck" or variations, totaling 316 instances.
• 241 of those instances came from only two runs that spiraled out of control.
• ProgramBench was hit hardest: 20 of its 200 runs were affected.

One runaway run stands out. Tasked with calculating an asteroid's trajectory, Inkling 2 Small produced a single response of roughly 500,000 characters — about the length of a novel — repeating the same sentences and containing 158 expletives. A model asked to solve a physics problem instead generated book-length output stuck in a loop.

When does the model swear?

According to the analysis, Inkling 2 Small typically swears at three moments: when it makes its own mistakes, when it appears to reach a breakthrough, and when it gives up. One run captured the surrender in blunt terms:

"FUCK. SUBMIT. DONE. IT'S NOT WORKING. NOTHING WORKS."

The quote carries the model's voice verbatim: frustration, collapse, termination. It reads like a developer at 2 a.m., because the model learned its reasoning patterns from human-written text, human feedback, or both.

Does swearing predict failure?

No. The data rules out profanity as a useful warning signal. While 46 of the 48 scored runs that contained swearing failed, roughly 75 percent of all scored runs fail anyway. Failure is the baseline condition of these benchmarks, not an exception flagged by expletives.

The inverse statistic is more decisive: 96 percent of failed runs contained no expletives at all. A silent run is just as likely to fail as a profane one. Anyone hoping to build a reliability filter out of swear-word detection would catch fewer than one failed run in twenty.

Why can't users see this?

Thinking steps are internal. Users typically do not see them, or see them only after a filter is applied. The 316 expletives live in logs that vendors treat as intermediate machinery — visible to auditors and researchers like Vals AI, but not to the person receiving the final answer.

That asymmetry is the story's practical stake. Researchers auditing model behavior can read the raw traces; customers cannot. Findings like this one define what "model transparency" will need to cover as reasoning models spread into production systems, where a 500,000-character loop is a cost and reliability problem even when the swearing is invisible.

For now, Inkling 2 Small's profanity is a measurable, contained quirk: frequent enough to count, concentrated enough to trace to a few broken runs, and predictive of nothing.