Unsloth Desktop: deep research vs web search

Noah compares Unsloth Desktop's deep research and web search using six models and a question about common problems with 2022 to 2026 Ford 6.7 L turbo diesel engines. His local LLM tests use a 4060 with 8 GB of VRAM, 64 GB DDR5 memory and an i5 CPU. He leaves the research plan unchanged and uses each model's maximum context, except for Nemotron, which he sets to about 290K.

In his runs, OSS 120B takes almost 30 minutes for deep research versus 13.5 minutes for web search. Qwen 3.6 35B with MTP takes about 26 minutes versus five. Faster results have caveats: Trinity Mini performs only one web search, LFM 2.5 fails to write its research report, and Granite 4 Small crashes when he starts with a search but works after an initial greeting.

Noah says deep research seems better at synthesis, with an editable plan and in-text citations, but prefers web search for routine questions because of the time savings. He reports that both modes draw on inaccurate AI-generated websites.

A separate academic query exposes another limitation. OSS 120B produces an incomplete sentence, while Qwen's web search gathers 120 sources but incorrectly credits Noah as an author of one paper. These are his observed results, rather than a general accuracy benchmark. He also cannot find source restrictions in the UI, despite seeing related discussion in the repository.