Qwen3.6-Plus repeats Gemma 4 research bugs

The sharp takeaway is that users do not care how ambitious the launch copy sounds if the agent loop feels sloppy in practice. If this anecdote reflects a broader pattern, Qwen’s latest release may be trading benchmark polish for worse tool discipline.

• –The official Qwen3.6-Plus launch positions it as an agentic coding model with a 1M-token context window and multimodal reasoning, so the expectation bar is high.
• –The post’s core complaint is not accuracy alone, but execution: repeated thinking blocks, search-only behavior, and failure to stop when enough evidence is already gathered.
• –That makes the comparison to Gemma 4 especially pointed, because both models are being judged on workflow quality rather than raw chat competence.
• –The user’s Qwen3.5 comparison suggests the regression is visible in real research tasks, where tool selection and persistence matter more than single-turn response quality.
• –One Reddit thread is not a benchmark, but it is the kind of anecdotal signal that shapes developer trust fast when a model is meant for agentic use.