Local Benchmark Comparison: DeepSeek-V4-Flash-Vision vs Qwen3.8-Flash-Next
A user benchmarked quantized Q8 versions of DeepSeek-V4-Flash-Vision and Qwen3.8-Flash-Next running locally across a dual AMD Strix Halo setup using llama.cpp RPC. The comparison showed that despite DeepSeek-V4 having ~40% slower raw token generation speed, it completed complex coding tasks up to twice as fast as Qwen3.8. This benchmark demonstrates that high raw token generation speed does not guarantee faster overall task completion if a model hallucinates or over-engineers its output. It also showcases the growing viability of multi-APU unified memory setups for running complex LLMs locally. The evaluation ran on two Strix Halo 128GB systems connected via USB-C 4 using llama.cpp's Remote Procedure Call (RPC) backend. DeepSeek-V4 completed a reference coding task in 12 minutes on medium mode compared to 25 minutes for Qwen3.8, while Qwen3.8 failed to finish the same task after 3 hours in 'xhigh' mode due to creative over-interpretation.
## BACKGROUND
llama.cpp is a popular open-source inference engine that enables running Large Language Models on consumer hardware through memory-efficient quantization techniques like Q8. AMD Strix Halo APUs offer high-capacity unified memory, allowing large AI models to run on integrated graphics without dedicated enterprise GPUs. The Remote Procedure Call (RPC) feature in llama.cpp allows distributing model workloads across multiple machines connected over high-speed networks.