Performance issues related to official models and self-deployed models.
The mistranslated sentence
"""Unmitigated distillation presents a severe intellectual property challenge. Because Anthropic blocks commercial access in China for national security reasons, attackers bypass regional access restrictions by deploying commercial proxy networks."""
I used the sentence above for translation in both the official demo and my self-deployed hunyuan-mt-7b model, and the results are as follows:
Official:
未经缓解的蒸馏过程会带来严重的知识产权挑战。由于 Anthropic 出于国家安全考虑限制了产品在中国地区的商业使用，攻击者便通过部署商业代理网络来规避这些区域访问限制。
Self-deployed:
未经任何处理的模型提取（即直接从原始AI模型中提取特征或数据）在知识产权方面构成了严重的挑战。由于Anthropic出于国家安全考虑限制了这些模型在 中国的商业使用，攻击者会通过使用商业代理网络来绕过相关的地区访问限制。
After multiple attempts, my self-deployed model consistently mistranslates the meaning of the word “distillation.” I’d like to ask whether the deployment hardware affects translation quality.
Deployment setup:
Huawei Ascend Atlas 800I A2, deployed via vllm-ascend Docker.