Qwen3.6-27B SimpleQA 95.7%

This is a fully local web-search agent combining a single RTX 3090 24GB, Ollama (qwen3.6:27b), and a langgraph_agent strategy based on LangChain create_agent().

• It used tool-calling, parallel subtopic decomposition, and up to 50 iterations.
• Grading was done via qwen3.6:27b self-graded method.
• Results were SimpleQA 95.7% (287/300) and xbench-DeepSearch 77.0% (77/100).
• The comparison table also presented Qwen3.5-9B 91.2% (182/200) and gpt-oss-20B 85.4% (295/346).
• These are agent + search scores, not closed-book performance.
• Caveats noted include possible SimpleQA contamination, LLM-judge noise, and no repeated re-testing; BrowseComp / GAIA figures are not yet available.
• It also added that xbench-DeepSearch is a Chinese-language benchmark, which may favor the Qwen family.
• It was compared as being at a similar level to Perplexity Deep Research (93.9%) and Tavily (93.3%).