A Reddit user shared an interesting test on Qwen-35B, a large language model (LLM). The experiment focused on the model's visual analysis and tool calling capabilities.

Test Details

The LLM was provided with a low-quality image and asked to locate a ring. Qwen-35B was able to analyze the image, understand the exact position of the ring, and, even more remarkably, use a Linux terminal to circle the corresponding area.

Performance

The user highlighted the model's processing speed, which reaches 100 tokens per second (tk/s) on consumer hardware, specifically a 3090 GPU. This suggests significant optimization for inference on less expensive hardware compared to enterprise solutions.