Local Deep Research hits 95.7% with Qwen3.6-27B

Hot take: this is more interesting as a systems result than a raw model win, because it suggests the agent/tooling stack is now doing as much work as the model size itself.

• –The headline number is strong, but the post is explicit that this is an agent + search benchmark, not a closed-book evaluation.
• –The setup is unusually practical: a single 3090, Ollama, multi-iteration tool calling, and parallel subtopic decomposition.
• –The claimed gain seems tied to tool-calling and orchestration quality, which matches the author’s hypothesis that newer Qwen models are better at agentic workflows.
• –The biggest caveats are real: possible SimpleQA contamination, self-graded judging noise, sample-size limits, and a Chinese-language benchmark advantage.
• –The release is compelling for local AI users because it combines privacy, encryption, reproducibility, and open-source distribution with competitive benchmark numbers.