Fable 5 and Mythos 5. This move immediately ignited a fascinating technological counter-response from Beijing.
Just two days later, a rival AI entity, Z.ai, began rolling out GLM-5.2. Crucially, this new model was reportedly trained entirely on domestic hardware—specifically Huawei Ascend chips—completely bypassing reliance on Nvidia silicon. This development signaled that the global race for cutting-edge AI is not just about software, but a fierce battle over foundational technology and supply chains.
Within a week of this digital standoff, GLM-5.2 quickly became a force, climbing to the top of openly available leaderboards and achieving remarkable performance metrics that put it squarely in competition with the U.S. giants. It rapidly escalated Z.ai’s market value, pushing its valuation past the one-trillion-dollar mark in Hong Kong.
The competitive landscape proved incredibly nuanced. While Fable 5 faced regulatory restrictions, GLM-5.2 demonstrated strong capability across various tests. In tasks focusing on human preference coding, GLM-5.2 edged ahead of its competitors. Furthermore, in multi-week knowledge tasks—where models test their ability to synthesize vast amounts of fragmented information—Fable 5 still led in one specific assessment, underscoring the complexity of measuring true intelligence.
However, when looking at raw performance and accessibility, the picture shifts dramatically. GLM-5.2’s open-weight nature means anyone can download, fine-tune, and self-host its weights, a stark contrast to closed systems. Despite this openness, the sheer scale of the model remains demanding: it requires substantial enterprise GPU clusters, emphasizing that creating frontier AI is less about code and more about infrastructure.
The hardware story adds another layer of complexity. Z.ai’s success relies on its proprietary stack, demonstrating that domestic silicon can indeed handle training-class jobs. Nevertheless, a significant gap remains between the performance of these specialized chips and the leading Nvidia technology. While reports suggest certain Chinese clusters can handle full-parameter post-training of models like DeepSeek’s V4, it underscores a critical point: model parity does not automatically equate to hardware parity.
The restrictions imposed by the U.S. government on advanced AI chips are designed to control access to high-end computing resources. Yet, as industry observers note, the potential for domestic silicon to close this gap is real. Meanwhile, the market is buzzing with anticipation. As investment lock-ups expire and whispers circulate about a Chinese equivalent to Fable 5 arriving sooner than expected, the stakes for the future of AI dominance are reaching a fever pitch.
Credit: Tom’s Hardware
