35B Local Powerhouse

Using Qwen3.5 35B, a paper-based webapp was built using only prompts and a step-by-step guide, and afterward that guide was converted into skills with added examples.

• Repository: research-webapp-skill
• Examples were also organized into a separate folder so they can be reused for other papers

The runtime environment was a 16GB VRAM RTX 5080 Laptop, and llama-server was used with the following settings.

• --jinja, --reasoning off
• -c 90000
• --context-shift, --keep 1024

Subjectively, this model maintained task context well even at 70-80K context, and on the same task, Gemma4 26B MoE failed to generate the app even with the latest Jinja template and an Unsloth model.

However, the author left open the possibility that they may have misconfigured something, and added that they are still using Gemma4 for other purposes.