Z.ai announced GLM-5.3-Flash on 26 August, confirming it was the model anonymously released under the name Ox Alpha since 20 August. Announced characteristics: mixture-of-experts architecture with 320 billion parameters, 18 billion of which are active, a one-million-token context, natively multimodal, released under the MIT licence. And one detail that goes beyond the technical: the lab says it runs entirely on Chinese AI chips.
We have been following this story since Monday, refusing to name a developer until a primary source did so. The primary source has spoken, and the revelation beats speculation.
What the clues had spotted correctly
Credit where credit is due to the community's work. Tokenizer analyses, error message formats and video token accounting all pointed towards the GLM family. That was correct.
What those clues could not determine was the version. A tokenizer is usually shared across an entire model family: it identifies a lab, not a specific model. That is why we stuck to a direction rather than a name.
The name, now confirmed, is GLM-5.3-Flash. And it must be distinguished from GLM-5.3, announced on 14 August, which we covered for its vulnerability detection capabilities: that one is a text model whose weights were announced for roughly two weeks after launch, pending a safety evaluation. These are two different models.
The announced characteristics
The architecture. A mixture of experts of around 320 billion parameters in total, with around 18 billion active per token, with 8 experts summoned out of 288. That is exactly the mechanism we described in our article on mixture of experts, and it explains the name: Flash here refers to inference efficiency, not a small size.
The context. One million tokens, with community tests indicating the capacity does indeed hold beyond 900,000 tokens.
Multimodality. Text, image and video input. The lab says it trained the model to inspect rendered interfaces, gameplay sequences and 3D outputs, then correct its own work based on that visual feedback. That is consistent with agentic usage where the model checks what it has produced.
The pricing. Around $0.15 per million input tokens and $0.50 per million output tokens, with $0.03 for cached input.
This is what no one had anticipated. The model is released under the MIT licence, one of the most permissive in existence, with published weights. We had hypothesised that a stealth release was meant to prepare a closed commercial launch. It is the opposite: the lab used anonymity to test, then opened everything up. This contrasts with Alibaba's Wan 3.0, pre-announced as open and ultimately released closed. Two Chinese labs, two opposite trajectories.
The point that goes beyond the technical
The lab states that the model runs entirely on Chinese AI chips. That sentence deserves a pause, because it is probably the most consequential piece of information in the whole affair.
For months we have been documenting the export controls designed to limit China's access to cutting-edge accelerators, and we have covered the rise of domestic alternatives. The demonstration here is public and verifiable through use: a 320-billion-parameter model, serving tens of millions of daily requests, on non-US hardware.
This comes at precisely the moment when Washington is reportedly considering asking dozens of countries to pick their technological camp. The message is hard to miss: the dependency they sought to create is being circumvented.
A customary caveat applies: this claim comes from the lab and is not audited. It is nonetheless consistent with the Chinese investments in domestic accelerators that we have documented.
The strategy, re-read in light of the outcome
Everything becomes clear in retrospect. In six days of anonymity and free access, the model reached around 70 million daily requests and embedded itself in the agent harnesses developers use daily.
A classic launch would have produced a press release, a benchmark table, and gradual adoption hampered by scepticism towards a Chinese lab. Here, thousands of developers judged the model on its results before knowing where it came from.
That is a remarkably effective demonstration of what we wrote on Monday: brand matters less than people think. And it settles the question in advance, since adoption preceded the reveal.
Points to verify
Two caveats, because not everything is consistent at this stage.
A discrepancy in specifications. An independent ranking currently lists the model with a 400,000-token context and describes it as proprietary, which contradicts the lab's announcement. This is likely a lag in updating or a measurement taken on a different access point, but caution is warranted on the figures until this is clarified.
Performance. The lab reports sharply improved results on its internal tests and a leading position on an independent index. As always, these measurements depend on the protocol used, and cross-lab comparisons remain delicate.
And the data sent during the preview?
The question remains open and deserves to be asked. For six days, thousands of developers ran entire code repositories through an anonymous provider whose page stated that requests are retained.
Knowing who was behind it changes nothing about what was transmitted. It simply makes it possible to identify who to contact, and under which legal framework. That is better than nothing, and it is far from a guarantee.
What to take away
This story will remain a textbook case, and for better reasons than the ones people thought on Monday.
A lab tested its model blind on real-world usage, let the product be judged on its own merits, then published everything under a permissive licence. Methodologically, that is more rigorous than most launches accompanied by in-house benchmark tables.
And the industrial message is clear. A leading model, trained and served on domestic hardware, released free for everyone, adopted en masse before it was even named: that is a response to chip access restrictions, and it is more eloquent than any diplomatic statement.