Qwen — what creators actually do with it

Across 5 posts, Qwen is the model creators reach for when they want an agent running on their own hardware. @gigaqian covers Qwen 3.8 27B, which quantizes to about 17 GB at 4-bit and fits on a single RTX 3090 or 4090, and advises using it as the brain of a local agent rather than for chat; the benchmark numbers in the post are Qwen's own, and the creator says to wait for independent evals (Qwen 3.8 27B). Unsloth is mentioned as already producing quantized versions.

The hands-on examples run smaller. @cyril.engman carries a rooted OnePlus 8 running Kali NetHunter and a local Qwen model, air-gapped, as an autonomous pen-testing agent he uses to harden his own networks, with every action audit logged (phone agent). @codrkai runs an abliterated 9B variant (a version with refusals stripped out) on a 32 GB MacBook Pro through llama.cpp and points an agent harness at it (local 9B). Both appear in the autonomous agents guide.

An older 2025 post introduces Qwen 2.5 Max as a free model covering chat, code, search, image and video, with the creator claiming it matches ChatGPT (China's open-source trio); no other post tests that. DeepSeek co-occurs in 2 posts. One entry, a DIY Claude Code phone remote, tags Qwen without using it (phone remote). The material here fits people with a 24 to 32 GB machine who want a local agent model, and it is thin on failures or limits beyond hardware size.