On the Limitations of Non-GPU AI Accelerators for Large-Model Inference: A Field Study of MoE and Multimodal Serving on Huawei Ascend

Benchmark Model Rank Results
code-generation-on-deepsweDeepSeek-V4-Flash-Vision-Exp (DeepSeek Harness)–Pass@1: 0.593
code-generation-on-nl2repoDeepSeek-V4-Flash-Vision-Exp–Accuracy: 57.7
code-generation-on-terminal-bench-2-1DeepSeek-V4-Flash-Vision-Exp (DeepSeek Harness)–Accuracy: 83.9
vulnerability-reproduction-on-cybergymDeepSeek-V4-Flash-Vision-Exp (DeepSeek Harness)–Success Rate: 75.3