316 Instances of 'Fuck' in Inkling 2 Small's Reasoning Traces, But Not a Failure Signal
Vals AI combed through 2,655 benchmark runs and found the word "fuck" and its variants a total of 316 times in the model's reasoning traces. Yet 96% of failed runs contained no profanity at all, which makes swearing a poor error indicator.
AI evaluation firm Vals AI said on November 6 local time that it had found profanity in Inkling 2 Small’s reasoning traces. Across 2,655 runs spanning 13 benchmarks, 50 runs contained the word “fuck” or a variant, for a total of 316 occurrences. The analysis landed one day after Thinking Machines Lab announced the model on November 5.
When the swearing showed up
The profanity was heavily concentrated in a handful of runs. Of the 316 total occurrences, 241 came from just two runs. According to Vals AI, the swearing generally appeared late in runs that involved a large number of steps, and those runs often ended in failure.
The material examined consisted of the reasoning traces and messages the model itself produced when run at default settings through the API, the interface developers use to wire a model into their own services. Tool calls and the results returned by those tools were excluded from the review. How often profanity turns up in the answers actual users see was not covered by this study.
Profanity is also a poor early warning for a failed answer. Vals AI reported that 96% of failed runs contained no profanity whatsoever. In other words, checking for swearing alone does not tell you whether an answer can be trusted.
What changed from the previous model
Inkling 2 Small is a second-generation model that accepts text and images and responds in text. It also lets you dial how deeply it thinks through a problem. Inkling-Small, released in July, accepted audio input, but the new model does not. Anyone working with audio material should confirm that difference first.
On the AA-Briefcase evaluation the company published, the new model scored 1,527. That puts it above GLM-5.3-Flash (1,451) in the same comparison and below DeepSeek-V4.1-Flash-1017 (1,570). The company noted that the new model sometimes fails to keep its answers concise, and sometimes fails to pick the right tool for a task. Teams considering it for production work should weigh those weaknesses alongside the scores.
Where you can try it
The company said the new model has been available to try in Tinker Playground since launch day. In Tinker, it can also be used to generate responses or to further train the model on your own data. The full model files, or weights, are slated for release next week under the Apache 2.0 license once external safety testing is complete.
In Vals AI’s analysis, profanity was not a signal that predicted failure. The open questions are how often swearing surfaces in answers delivered to real users, and why. Korean users will also want to check availability in Korea, Korean-language performance, and pricing before adopting it.