When a model generates tokens more slowly yet finishes tasks in half the time, the usual local inference benchmarks need a second look. That is the case in a comparison run on self-hosted hardware: two StrixHalo nodes with 128 GB each, connected through USB-C 4, using Llama (RPC) as the inference engine. The models tested are DeepSeek-V4-Flash-Vision and Qwen3.8-Flash-Next, both with Q8_K_XL quantization.

The most interesting figure is not generation speed. DeepSeek-V4-Flash-Vision is about 40 percent slower in raw token production, yet completes the same tasks in about half the time. The reason lies in the different behavior of the two systems. Qwen3.8-Flash-Next tends to overinterpret instructions: when the prompt is not highly detailed and leaves room for creative interpretation, the model adds unrequested elements and gets stuck in the details. According to the field observation, DeepSeek-V4-Flash-Vision produces fewer hallucinations and encounters fewer obstacles along the way.

The gap is stark in more aggressive modes. A task that Qwen3.8-Flash-Next completed in 25 minutes in 'medium' mode was not completed in about three hours in 'xhigh' mode. DeepSeek-V4-Flash-Vision finished the same task in 12 minutes in the fastest run in 'medium', and in 37 and 44 minutes in the two 'max' iterations.

For teams running professional workloads on their own infrastructure, the comparison challenges the raw tokens-per-second metric. A model that generates output quickly but produces text to be corrected, reworded, or discarded shifts cost from raw compute to human intervention. The TCO of an on-premise deployment is therefore not just about peak speed, but about the number of iterations required to get a usable result. In this scenario, DeepSeek-V4-Flash-Vision's lower tendency to drift into side interpretations becomes an operational advantage even with slower tokens.

To be clear, the test does not undermine Qwen overall. The author reports being a big fan and heavy user of both Qwen 3.6 27B and Qwen 3.8 27B. The point is more specific: once the hardware can run DeepSeek-V4-Flash-Vision, the preference for professional code shifts. And so does the criterion for evaluating local models: less focus on the brilliance of a single generation, more focus on the real time needed to complete a task without friction.