Two announcements, ten days apart, tell the story of where AI is heading next. On July 16, Moonshot AI released Kimi K3, a 2.8-trillion-parameter model whose Mixture-of-Experts design and Kimi Delta Attention architecture let it rival top American systems while activating only a sliver of its parameters per token.
Event Context
A month earlier, on June 13, Beijing-based Z.ai — the company formerly known as Zhipu — released GLM-5.2.
Both companies upended Anthropic and OpenAI’s narrative. They showed that China was capable of building frontier_level models and that the actual weights can be downloadable and runnable on hardware locally.
That single choice is now the clearest line dividing the two camps building the frontier.
Z.ai’s GLM-5.2 climbed to the top of the Artificial Analysis Intelligence Index among open-weight systems, scoring within striking distance of Claude Opus 4.8 and GPT-5.5 on demanding coding benchmarks like SWE-bench Pro and FrontierSWE, while costing roughly a fifth to a seventh as much to run.
Moonshot’s Kimi K3 — a hosted API and consumer app first — was released on July 16, with the full weights held back until Monday. That promise was kept: Moonshot published K3’s weights on July 27 under a Modified MIT license, making the 2.8-trillion-parameter model available for anyone to download, fine-tune and self-host.
Founder Yang Zhilin has framed the openness itself as the strategy as he wants to grow Moonshot’s user base through broader availability than competing US proprietary systems offer.
The closed camp has taken the opposite bet to keep the weights, sell access, and protect the model as the product, and they still have a real edge on raw capability as independent trackers put Fable 5, Opus 4.8 and GPT-5.5 ahead of GLM-5.2 and Kimi K3 on the toughest reasoning benchmarks, and some estimate the gap between the best closed and best open systems at roughly seven months.
Player Focus
Chinese-made accelerators still trail Nvidia’s Blackwell chips on performance per watt, so matching US compute gigawatt-for-gigawatt takes more chips and more floor space. But “trailing” is a very different problem than “cut off,” and Z.ai has now shown the gap is one you can build your way around.
And the chip story doesn’t end there as Hefei-based ChangXin Memory Technologies (CXMT) has been racing to close the gap.
CXMT is now China’s largest DRAM maker, and is targeting mass production of domestic HBM3 by the end of 2026. The company has reportedly supplied HBM samples to Huawei, whose Ascend accelerators are the same chips powering Z.ai’s new data centre.
Z.ai’s gigawatt facility proves China can already assemble a chip at scale; CXMT is the bet that China can also supply the memory that chip needs to be competitive, closing the last major import dependency in the domestic AI stack.
But gaps still remain as CXMT still needs access to the world’s most advanced chipmaking tools remains, which are restricted.
So does this become the smartphone story all over again — Chinese AI as Android, cheap, everywhere, running on hardware nobody controls centrally, versus a smaller number of premium, tightly integrated American systems sold the way Apple sells iPhones?
That analogy holds in places. Open weights plus cheap API pricing plus a domestic chip supply chain is close to the Android playbook. They give away the core software, win on distribution and volume, let a fragmented ecosystem of distillers and fine-tuners do the rest of the work for free.
And the closed labs are behaving somewhat like Apple — arguing that vertical integration, safety review and tightly controlled deployment justify a premium, even as that same control has now made them a target for their own government’s export policy.
Where the analogy strains is control. Android succeeded partly because Google still owned the platform underneath the fragmentation. It isn’t obvious anyone owns the open-weight AI ecosystem the same way — GLM-5.2 and Kimi K3 compete with each other as much as they compete with Claude or GPT, and once weights are public, no single company can fully steer what gets built on top of them.
That looks more like a market splitting into a commodity tier and a premium tier at the same time, with the boundary between them determined less by capability and more by who’s willing to pay for the difference.
What does look durable is the emerging split in how AI gets consumed. Routine, high-volume tasks will increasingly be routed to whichever open, cheap model is “good enough,” while frontier reasoning work will stay on closed, premium models.
Team Analysis
Now that the weights are out, the pattern that followed Llama, Qwen and DeepSeek could well repeat as outside developers will distil, quantize and fine-tune the model within days, spinning up a whole downstream ecosystem Moonshot doesn’t control and doesn’t need to.
Match Outlook
Neither Z.ai nor Moonshot is doing this out of generosity. But giving weights away is a smart distribution strategy as it could potentially turn developers, cloud providers and enterprise IT units into resellers, and make it hard for anyone to under-price you as the base model itself is free.
source
In practice, “open” comes with real friction as the weights are not conducive for self-hosting unless the user is well-resourced company. Still, the move to make these AI models open is an intriguing move by the Chinese lab.
Beyond the models, there is a more structural shift. According to a Bloomberg report Z.ai has completed and partially activated a 1 GW data centre, built entirely on Chinese-made chips, with clusters of more than 10,000 domestic accelerators. Zero Nvidia silicon inside.
Z.ai has said its recent GLM models were trained on Huawei’s Ascend accelerators running Huawei’s own MindSpore software stack, which the company describes as the first major open model built on a fully domestic hardware and software chain.
The company didn’t have a choice as the US placed it on its export blacklist in early 2025, cutting off legal channels to Nvidia’s advanced chips.
Now, this new data centre is China’s response to having no other option. And Z.ai’s move comes on the heels of Beijing drafting a roughly $295 billion, five-year national plan to build AI data centres with at least 80% domestic sourcing.
