Issue 2026-11-07 · Industry · 研究 · 模型发布
Thinking Machines model swore 316 times in benchmark runs

An analysis by Vals AI found Thinking Machines' Inkling 2 Small swore during internal reasoning 316 times across 2,655 benchmark runs. Profanity usually appeared late in long runs that failed anyway, but 96 percent of failed runs had no cursing, so it is not a useful warning signal.
Read original ↗- Boson AI Releases 45B Full-Duplex Model Higgs Realtime Duplex
- Inkling 2 Small launches as 594B-parameter MoE model
- Study: Chinese AI Models Politically Complex, Not Monolithic
- Vals AI: DeepSeek V4.1 Flash beats GPT 6 Sol on cybersecurity task
- Research: AI agent teams cost up to 5.1x more for tiny quality gains