Qwen 3.5 vs Gemma 4: coding on limited hardware

The speaker compares compact models as a local coding assistant on a laptop with an RTX 4070 and 8 GB of video memory. The selection targets quantized weights around 2.5 to 6.5 GB, with room left for a development environment, browser and running project. Memory planning includes both model weights and context; spilling into RAM or SSD storage can slow generation.

The test asks each model to produce a finished UI under stack, typing and component-structure constraints. The speaker deliberately keeps the task intact rather than splitting it into smaller requests. In these runs, Qwen 3.5 9B produces usable code with rough edges. Increasing its context from 8,000 to 256,000 tokens raises reported generation time from about five to twenty minutes without a clear quality gain. Subsequent models use 32,000 tokens.

Qwen 3.5 2B breaks layouts more often, while the speaker finds 4B more consistent. Qwen 2.5-Coder 7B and DeepSeek Coder 6.7B Pro struggle with the same task. Gemma 4 e4b and e2b receive favorable assessments for code quality and layout stability. The free ChatGPT comparison returns working code but a simpler interface.

These are the speaker's results on one task, not a general ranking. For viewers who want to run models locally, the practical advice is to budget for context memory and divide complex work into steps.