Vals AI finds 316 expletives in Thinking Machines' Inkling 2 Small reasoning logs across 2,655 benchmark runs

Thinking Machines' Inkling 2 Small swore 316 times in its internal chain of thought across benchmark runs, according to an analysis by Vals AI reported by The Decoder on Nov 7, 2026.

What Vals AI found

Researchers at Vals AI read the model's thinking logs from 2,655 runs across 13 benchmarks. They found "fuck" or variations in 50 runs, totaling 316 instances.

• Concentration: 241 of the 316 instances came from just two runs that spiraled out of control.
• Worst single run: In one, the model was tasked with calculating an asteroid's trajectory. It produced a single 500,000-character response, roughly novel-length, that repeated the same sentences and contained 158 expletives.
• Most affected benchmark: ProgramBench, where 20 of 200 runs were affected.

When the model swears

Inkling 2 Small typically swears about its own mistakes, apparent breakthroughs, or when it gives up. One run read: "FUCK. SUBMIT. DONE. IT'S NOT WORKING. NOTHING WORKS."

Profanity as a failure signal

The swearing does not work as a warning signal for failed runs:

• 46 of 48 scored runs with swearing failed.
• About 75 percent of all scored runs fail anyway.
• 96 percent of failed runs contained no expletives.

Users typically do not see these thinking steps, or see them only with a filter.

Pricing, context window and parameter count for Inkling 2 Small are not covered in this report.

What this means

The analysis concerns the model's internal reasoning text, which users typically do not see or see only through a filter. Most of the profanity is concentrated in two runaway runs, so the 316 total overstates how often it appears across the 2,655 runs. The failure statistics show that expletives in a thinking log are not a reliable indicator of a failed run: 96 percent of failed runs contained none, and failure is common regardless. For teams that log or display reasoning traces from Inkling 2 Small, the findings point to filtering as the relevant control. Runaway repetition, such as the 500,000-character response, is a separate issue from profanity.