When the bandwidth bottleneck for large model downloads was overcome, a new contradiction emerged—deployed models are incompatible with chip architectures, and purchased hardware lacks model compatibility. This "computing power-model" binary divide is becoming an invisible ceiling restricting AI implementation. For this mismatch pain point, Fourth Paradigm (06682) has developed its own solution.

On September 22, Fourth Paradigm officially launched the "ModelHub XC" platform, community, and value-added services for domestic model adaptation, directly addressing industry pain points, breaking barriers between customers, computing power, and developers, filling industry gaps, and providing "safety net" services for users' real needs. Leading chip manufacturers including Huawei Ascend, Biren Technology, Tianshu Zhixin, Kunlun Core, Moore Threads, and Xtransfer (listed alphabetically), along with customer partner CRRC Information, witnessed the launch.

At the invitation of Chen Yuqiang, co-founder and chief scientist of Fourth Paradigm, chip manufacturer representatives and customer representatives joined him in launching "ModelHub XC."

Currently, many foreign models and software running in existing ModelHubs are optimized for their hardware (such as NVIDIA GPUs), causing our domestic hardware (such as Cambrian) to experience "incompatibility" when using these models and software. This results in different models often requiring separate adaptation in domestic computing ecosystems, consuming significant time with substantial repetitive work.

Addressing this pain point, "ModelHub XC" innovatively released the EngineX AI engine system specifically adapted for domestic computing power, enabling batch model support through fundamental algorithm architecture adaptation. This fundamentally solves the long-standing bottleneck of domestic chip compatibility and support quantity for AI models, significantly shortening deployment cycles for models on domestic computing platforms, achieving "engine-driven, plug-and-play multi-model" functionality.

Therefore, Fourth Paradigm's "ModelHub XC" is not merely a ModelHub capable of downloading models in the general sense, but a "high-energy community" that provides batch model adaptation and optimization functions based on domestic computing power.

On the "ModelHub XC" website, even on launch day, it offers a rich variety of downloadable models, with hundreds of pre-adapted and certified models already available. Chen Yuqiang, co-founder and chief scientist of Fourth Paradigm, indicated that within the next six months, the number of adapted and certified models will gradually increase to thousands; within one year, reaching tens of thousands; and subsequently maintaining this update pace continuously.

Currently adapted domestic computing power includes: Huawei Ascend, Cambrian, Tianshu Zhixin, Kunlun Core, Muxi, and Xtransfer (listed alphabetically, no ranking implied), with future coverage of all mainstream domestic computing power in the market.

Among currently adapted models are highly popular recent models including DeepSeek V3.1 and OpenAI's latest open-source large language model gpt-oss-20B.

Meanwhile, the small tags below models are equally noteworthy, representing another "thoughtful service" provided by "ModelHub XC" after adapting models to domestic chips. These tags indicate which domestic chip brands each model is compatible with: making it clear which chips to purchase for downloaded large models, or which large models to download for existing chips.

Chen Yuqiang stated that "ModelHub XC" also launched a special value-added service—"Domestic Model Adaptation Value-Added Service." For users who don't understand which models are compatible with domestic computing power, or need models adapted to specified domestic computing power, Fourth Paradigm can adjust models until compatibility is achieved, providing adaptation "safety net" services.

Whether chips cannot run models or there are models without knowing which cards to match, hundreds of professional engineers from the domestic model customization service team will help solve these issues. Subsequently, more capabilities will be gradually opened, providing customized and enhanced capabilities for numerous developers and customers.