Developer questions need for multilingual weights in dense coding models like Qwen

A developer running Qwen3.6-35B-A3B and the dense Qwen3.6-27B argues that multilingual weights waste VRAM for users focused solely on English technical text and code.

• The author notes that dense models require loading full 16-20GB weight files, locking out users with limited hardware like 8GB or 12GB GPUs.
• They propose two potential solutions: training new models from scratch with zero multilingual data, or pruning existing models to remove non-English capacity post-hoc.
• The author references a study on reasoning models like DeepSeek-R1 and OpenAI o1, noting their low execution accuracy on complex tasks compared to simpler benchmarks.

The core concern is whether it is technically feasible to strip multilingual weights to reclaim VRAM for better coding performance on consumer hardware.