Qwen3.6 vision stalls in llama.cpp

This is not a launch story, but it is the kind of deployment friction that decides whether a strong open multimodal model actually gets used locally.

• –Qwen’s official materials position Qwen3.6-35B-A3B as a 35B-total, 3B-active open MoE model with strong multimodal and coding performance, so slow image prefill in common GGUF workflows matters.
• –The user’s text-token speed looks healthy at around 65 t/s, suggesting the bottleneck is likely vision preprocessing, projector placement, llama.cpp integration, or model-specific multimodal overhead rather than normal decoding.
• –Gemma4’s much faster image startup on the same machine is the important comparison: local multimodal users care about time-to-first-token after image upload, not just tokens per second after generation starts.
• –The thread is still anecdotal and small, so this is a signal to watch rather than proof of a broad Qwen3.6 regression.