Qwen3 has a major problem

Alibaba has introduced Qwen3, a new open-weight family of models featuring two MoE (Mixture-of-Experts) variants and six dense models. According to Alibaba, the top model of Qwen3 outperforms competitors like OpenAI's o1, DeepSeek-R1, and Gemini 2.5 Pro. However, real-world testing presents mixed results. Ivan tested Qwen3’s 30B and 32B models with identical prompts, and the models passed the hexagon and strawberry tests, as well as generating a functional game prototype. Despite these successes, Qwen3 exhibited some significant issues. One notable problem, identified by Theo, is the model's tendency to overthink. In some cases, it consumed thousands of tokens simply processing its response, exhausting its context before delivering an answer. In one instance, it repeatedly said "wait" 32 times during a code prompt. This issue intensified on Qwen's official chat site, where the model said "wait" 81 times, though a workaround involving the /no\_think prompt was found to resolve the issue.