REDDIT · REDDIT// 2h agoTUTORIAL

llama.cpp vision needs mmproj, multimodal CLI

This Reddit thread is a practical guide to getting Qwen3.5-4B GGUF vision working in llama.cpp. The poster found that the separate mmproj projector and the multimodal server path work, while plain llama-cli did not.

// ANALYSIS

The key correction is that llama.cpp multimodal support is not exposed through plain llama-cli; the working path uses the multimodal CLI or llama-server with a separate mmproj file. The post is useful as a real-world troubleshooting note, and the 20 tokens/sec question reads more like a configuration and benchmarking issue than a model limitation.

// TAGS

llama-cppqwen3.5multimodalvisionggufmmprojlocal-llminferenceperformance

DISCOVERED

2h ago

2026-04-20

PUBLISHED

4h ago

2026-04-20

RELEVANCE

7/ 10

AUTHOR

Dabber43