When the bandwidth bottleneck for downloading large models is broken, a new contradiction emerges—the deployed models are incompatible with the chip architecture, and the purchased hardware lacks model adaptation. This binary split of &#34;computing power - model&#34; is becoming an invisible ceiling that restricts the implementation of AI. For this mismatch pain point, FOURTH PARADIGM (06682) has developed its own solution.
<img src="https://imageproxy.pbkrs.com/https://img.zhitongcaijing.com/image/20250922/1758540342972149.png?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="image.png" original-src="https://imageproxy.pbkrs.com/https://img.zhitongcaijing.com/image/20250922/1758540342972149.png"/>
According to Zhitong Caijing APP, on September 22, FOURTH PARADIGM officially launched the &#34;Xinchang Model Box&#34; ModelHub XC platform, Xinchang community, and Xinchang model adaptation value-added services, directly addressing industry pain points, breaking down barriers between customers, computing power, and developers, filling industry gaps, and providing &#34;safety net&#34; services for users' real needs.
<img src="https://imageproxy.pbkrs.com/https://img.zhitongcaijing.com/image/20250922/1758540360530189.png?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="image.png" original-src="https://imageproxy.pbkrs.com/https://img.zhitongcaijing.com/image/20250922/1758540360530189.png"/>
Leading chip manufacturers such as Huawei Ascend, Biren Technology, Tensilica, Kunlun Core, Moore Threads, and Horizon witnessed the event alongside customer partner CRRC Information. At the invitation of Chen Yuqiang, co-founder and chief scientist of FOURTH PARADIGM, representatives from chip manufacturers and customers jointly launched the &#34;Xinchang Model Box&#34; ModelHub XC.
<img src="https://imageproxy.pbkrs.com/https://img.zhitongcaijing.com/image/20250922/1758540375540353.png?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="image.png" original-src="https://imageproxy.pbkrs.com/https://img.zhitongcaijing.com/image/20250922/1758540375540353.png"/>
Currently, many foreign models and software running on ModelHub in the market are optimized for their hardware (such as NVIDIA GPUs), making it difficult for our domestic hardware (such as Cambricon) to use these models and software. This leads to the need for different models to be adapted separately in the domestic computing power ecosystem, resulting in long time consumption and a large amount of repetitive work.
<img src="https://imageproxy.pbkrs.com/https://img.zhitongcaijing.com/image/20250922/1758540387443436.png?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="image.png" original-src="https://imageproxy.pbkrs.com/https://img.zhitongcaijing.com/image/20250922/1758540387443436.png"/>
To address this pain point, the &#34;Xinchang Model Box&#34; has innovatively launched an AI engine system EngineX specifically adapted for Xinchang computing power, driving batch model support through basic algorithm architecture adaptation. This fundamentally solves the long-standing bottleneck of AI model compatibility and support quantity for domestic Xinchang chips, significantly shortening the deployment cycle of models on domestic computing power platforms, achieving &#34;engine-driven,
&#34;Multi-model plug-and-play.&#34; Therefore, the &#34;Xinchuan Model Box&#34; released by FOURTH PARADIGM is not just a ModelHub that can download models in a general sense, but also a &#34;high-energy community&#34; that provides bulk model adaptation and optimization functions based on Xinchuan computing power.
<img src="https://imageproxy.pbkrs.com/https://img.zhitongcaijing.com/image/20250922/1758540402718508.png?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="image.png" original-src="https://imageproxy.pbkrs.com/https://img.zhitongcaijing.com/image/20250922/1758540402718508.png"/>
<img src="https://imageproxy.pbkrs.com/https://img.zhitongcaijing.com/image/20250922/1758540411275578.png?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="image.png" original-src="https://imageproxy.pbkrs.com/https://img.zhitongcaijing.com/image/20250922/1758540411275578.png"/>
On the &#34;Xinchuan Model Box&#34; webpage, it can be seen that on the day of its launch, it offered a rich variety of downloadable models, with the number of models that have been adapted and certified already in the hundreds. Chen Yuqiang, co-founder and chief scientist of FOURTH PARADIGM, introduced that in the next six months, the number of adapted and certified models will gradually be updated to the thousands; within a year, it will reach the tens of thousands; and this update speed will be continuously maintained. The Xinchuan computing power that the platform has currently adapted includes: Huawei Ascend, Cambricon, Tianzu Zhixin, Kunlun Core, Muxi, and Xiwang, etc. (sorted by English name, no particular order), and in the future, it will cover all mainstream Xinchuan computing power on the market.
Among the models that have already been adapted, there are several recently highly regarded large models such as DeepSeek V3.1 and OpenAI's latest open-source large language model gpt-oss-20B, which are &#34;popular contenders.&#34;
At the same time, the small labels below the models are also noteworthy. They represent another &#34;thoughtful service&#34; provided by the &#34;Xinchuan Model Box&#34; after adapting the models to domestic Xinchuan chips. The labels indicate which domestic chip brands the model can adapt to: it is clear which chip to purchase after downloading the large model, or which large model can be used with the chip.
<img src="https://imageproxy.pbkrs.com/https://img.zhitongcaijing.com/image/20250922/1758540431784859.png?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="image.png" original-src="https://imageproxy.pbkrs.com/https://img.zhitongcaijing.com/image/20250922/1758540431784859.png"/>
Chen Yuqiang stated that the &#34;Xinchuan Model Box&#34; has also specially launched a value-added service - &#34;Xinchuan Model Adaptation Value-Added Service,&#34; aimed at users who are unfamiliar with which models are adapted to domestic computing power and need models adapted to specific domestic computing power. FOURTH PARADIGM can adjust the models until they are adapted, achieving a &#34;safety net&#34; for adaptation. Whether the model cannot run on the card, or if there is a model but one does not know which card to match, the hundreds of professional engineers in the Xinchuan model customization service team will help you solve the problem. In the future, more capabilities will be gradually opened up to provide customized and stronger capabilities for many developers and customers

06682

On September 22nd, FOURTH PARADIGM launched the "ModelHub XC" platform, aimed at addressing compatibility issues between domestic GPUs and large models. This platform provides value-added services for the Xinchuang community and model adaptation, breaking down barriers between customers, computing power, and developers, filling a gap in the industry. Through the AI engine system EngineX, specifically adapted for Xinchuang computing power, the platform significantly shortens the deployment cycle of models on domestic computing power platforms, achieving plug-and-play functionality for multiple models

Fill the gap! FOURTH PARADIGM releases "Xinchuan Model Box" ModelHub XC, connecting domestic GPUs and domestic large models