APMIC (Accelerate Private Machine Intelligence Company), a leading provider of enterprise AI solutions, announced today (August 8th) that its PrivModel service fully supports OpenAI’s recently launchedGPT-OSS series models, achieving an approximately 40% improvement in accuracy and a 90% reduction in inference costs in specific application scenarios. APMIC also appointed Ian Chen, a veteran GPU and AI chip expert, as General Manager of the United States, further strengthening local operations and strategic collaboration in the North American market and promoting the global development of private AI initiatives.
GPT-OSS combined with PrivModel: high accuracy and low cost
APMIC has fully upgraded its original "S1 Distillation" solution to PrivModel, integrating it deeply with the OpenAI gpt-oss series (including gpt-oss-120b and gpt-oss-20b). Internal testing has shown that through fine-tuning and distillation techniques, the model maintains stable performance at FP4 precision, improving accuracy by an average of 40% and significantly reducing inference costs by 90%.
The distilled model can maintain efficient operation on the NVIDIA RTX Pro 6000 platform and supports Traditional Chinese Chain of Thought (CoT) reasoning for the first time, allowing enterprises to privately deploy "Reasoning AI" within a regulatory framework.
APMIC noted that traditional QLoRA (Quantized LoRA, large language model fine-tuning) methods still have limitations in learning specialized knowledge. Therefore, the team combined CPT, SFT, and RL methods to achieve deep fine-tuning. In the TMMLU+ local knowledge test, all GPT-OSS models achieved scores above 80, a significant improvement over the 2023 GPT-3 score of less than 60. In the medical application field, they even outperformed the O4 Mini by 13 points, while their mathematical logic performance was only 2 points lower.
In addition, PrivModel complements the shortcomings of the gpt-oss series in visual capabilities, helping companies build exclusive and commercially valuable private AI models.
Deeply cooperate with industry chain partners to ensure safe implementation
APMIC CEO Jerry Wu said that when paired with the "PrivStation" architecture, PrivModel can provide operational performance comparable to the o4 mini level. With the support of computing partners such as CKmates, GMI Cloud, Glows AI, and Ubilink, as well as hardware manufacturers such as Dell, HPE, and NVIDIA, the model can run stably on the latest Blackwell display architecture devices.
At the same time, APMIC also combines CyCraft's security protection to ensure that enterprises can not only effectively lock in their own knowledge assets during the deployment of private AI, but also enjoy high-efficiency inference at a lower cost, truly realizing the application path from "model to decision".
Senior GPU experts lead the way in promoting the North American market
APMIC also announced the appointment of Chen Yinjun as General Manager of the US Market. Chen Yinjun has over 20 years of experience in the GPU and AI chip industry, having held engineering and business management positions at NVIDIA in Silicon Valley, Hong Kong, and Taiwan, specializing in high-end graphics cards and data center GPUs.
In the future, Chen Yinjun will focus on promoting the implementation of PrivModel and PrivStation in the North American market, and cooperate with GPU cloud services, OEMs, and design software companies to expand industrial applications such as finance and manufacturing that have high demand for compliant AI.
Through technological innovation and global deployment, APMIC hopes to enable enterprises to deploy AI at a lower cost, with higher accuracy and greater security, accelerate the industrial implementation of "privatized AI", and promote a comprehensive transformation from data protection to commercial value.
