Open Source

Unsloth user drops Qwen 3.6 mmproj to reclaim VRAM without harming text performance

Removing vision module from Qwen 3.6 saves VRAM while preserving text reasoning.

Deep Dive

A Reddit user removed the mmproj file from Qwen 3.6-35B-A3B by Unsloth to remove vision and save VRAM, then asked if this affects the model's text ability. The user runs it for agentic coding.

Key Points
  • Removing the mmproj file from Qwen 3.6-35B-A3B frees VRAM by eliminating the vision projection module.
  • The model by Unsloth is a mixture-of-experts (MoE) architecture popular for efficient local inference.
  • User reports zero text degradation for agentic coding tasks, though no rigorous benchmarks were shared.

Why It Matters

A simple file deletion lets developers run large multimodal models on consumer GPUs while keeping text capabilities intact.