Bijan Bowen compares two compressed Bonsai 2 variants of Qwen 3.8 27B, highlighting their much smaller weight files and optional memory-saving settings. He distinguishes broad retained-intelligence claims from weaker coding-specific results and notes that KV cache and vision components still affect total memory use.
Bijan Bowen's browser-OS test repeatedly stalls in rewrites, while a simpler PC-repair website prompt produces usable pages. His C++ and subway-game experiments expose a gap between compiling or displaying a scene and delivering functioning gameplay; later repairs by another model are not credited as Bonsai successes.
Bijan Bowen demonstrates that both variants can invoke Blender and produce simple city scenes, although one scene has floating buildings. Watch-site outputs look respectable but omit the requested 3D models, reinforcing the difference between aesthetic output and complete functionality.
Bijan Bowen gets more encouraging SVG and simple skateboard-game results at medium reasoning effort, with simplified prompts and targeted feedback. His comparison with the full model is exploratory rather than a controlled benchmark, and he concludes that the tiny footprint is impressive without equating these models to frontier systems.
Watch on YouTube
