Qwen 3.6 27B vs Gemma 4 31B: Tauri Markdown App Build-Off

AI Summaries (evaluation playlist)2026-05-13go watch the original →

the gist

Qwen 3.6 27B took 46min to generate a functional Tauri markdown viewer/editor needing minor port/server and Tauri method fixes; Gemma 4 31B did it in 20min after adding filesystem plugins, but skipped toolbar buttons.

The Breakthrough

Author tasked Qwen 3.6 27B dense and Gemma 4 31B dense models with building identical cross-platform Tauri markdown viewer/editors from the same prompt via OpenCode, revealing both produce working apps after brief manual fixes despite autonomous full-project implementation.

What Actually Worked

Models first analyze app description (two-panel viewer/editor prioritizing viewing with edit/save/open features, Tauri stack) and output phased plans with subtasks into separate Markdown files.
Run tauri init for project scaffolding, then instruct model to implement entire plan autonomously, self-correcting errors en route.
Qwen 3.6 generates code with toolbar, split editor/preview panes supporting real-time Markdown rendering and file open/edit; fixes needed: add frontend server start block (e.g., missing serve lines), update deprecated Tauri method name from v1.
Gemma 4 auto-organizes description/plan files into documentation/ folder; produces similar two-panel UI with edit/preview toggle buttons and file open; fix needed: add fs and related plugins to tauri.conf.json for filesystem access.

Before / After

Qwen 3.6 planning: 4 minutes; full implementation: 46 minutes. Gemma 4 planning: 2.5 minutes; full implementation: 20 minutes. Both yield launchable apps with core viewing/editing/preview after 2-3 line manual fixes each; no formal accuracy metrics.

Context

Author seeks daily local coding LLM, prefers dense over MoE architectures for code gen, tests on personal hardware (desktop GPU with ample VRAM, accessed via LAN from MacBook) using real task over benchmarks: Markdown viewer/editor unmet need. Stress-tests long-session autonomy; notes Gemma halves time but Qwen plans more granularly (twice phases/tasks); power draw spikes under load, impacting bills for constant use. Both viable, author plans dual-use pending further trials.

Notable Quotes

"for tasks related to writing code dense models tend to deliver better results"
"the model identifies its own mistakes during the process and thinks about how to fix them"
"Gemma listed all the tasks it worked on at the end of its response"
"both models delivered very similar results"

Content References

None explicitly cited beyond tools used.