GLM-5.3-Flash is the newest open-weight foundation model from Z.ai, the Beijing-based lab formerly known as Zhipu AI. Z.ai has launched its latest open-weight model, GLM-5.3-Flash, previously code-named Ox Alpha, saying that the system ran entirely on a cluster of 100,000 domestically produced chips during a high-profile stealth trial. The release is a 320B parameter mixture-of-experts model with 18B active parameters, natively multimodal, with a 1M-token context window, published under the MIT License so teams can self-host, fine-tune, and deploy without commercial restrictions.
The launch matters because it puts a name to a model that developers had been probing anonymously for over a week. Ox Alpha, the anonymous "stealth" model that showed up on OpenCode and OpenRouter with no branding attached, is GLM-5.3-Flash, the first natively multimodal entry in the GLM-5 line. During that preview period, the lab behind the model was able to serve 100 trillion free tokens per day , and Z.ai is now emphasizing that the whole stack ran on Chinese AI accelerators rather than Nvidia hardware, a point aimed squarely at Chinese enterprises navigating export controls and at global developers weighing price and availability.
Z.ai is the international brand for a company with deep academic roots. Founders Tang Jie and Li Juanzi, both professors at Tsinghua University's Department of Computer Science and Technology and active participants in the university's Knowledge Engineering Group (KEG), established Zhipu AI in 2019 at Tsinghua University Science Park in Beijing. For founders and operators building on open models, GLM-5.3-Flash gives them a permissively licensed, long-context multimodal option positioned to compete on cost against closed frontier APIs, backed by a lab that has now demonstrated it can train and serve at frontier scale on domestic silicon.
Comments (7)
Sign in to join the discussion.
N
Nadia B.Aug 27, 2026
320B-A18B under MIT is the actual headline and it's buried in bullet three. Whoever wrote this tweet needs a raise and also a demotion.
P
Priya F.Aug 27, 2026
Ox Alpha was the codename? You had the coolest preview brand in the game and swapped it for GLM-5.3-Flash. Marketing crime.
K
Kwesi OtooAug 27, 2026
Running entirely on Chinese chips is the sentence the whole industry is going to be quoting tomorrow. Curious if you plan to publish throughput numbers or leave that for the reverse engineers.
M
Marcelo D.V.Aug 27, 2026
Every model this year is 'natively multimodal with 1M context' at this point. What does natively actually mean when half the labs use the word like seasoning?
T
Tove H.Aug 27, 2026
A CFO reading 'highly competitive price' with no numbers attached is like a cat reading a menu. Give us the per-million-token figure or we assume the worst.
D
deviAug 27, 2026
Any word on rate limits for the free tier and whether the 1M context is actually usable or one of those 'technically supported' situations that OOMs at 200k?
B
bogdan k.Aug 27, 2026
MIT license on a 320B model is a genuinely wild move and I don't think people have processed it yet. The fine-tune community is about to eat very well.
Work With Us
Book an intro call to explore our suite of services. We might have a waitlist.
Z.ai
Inspiring the development of AGI to benefit humanity.
Inspiring the development of AGI to benefit humanity.
Tags
TLVC Rating
Community Rating
No ratings yet
to rate this launch.
GLM-5.3-Flash is the newest open-weight foundation model from Z.ai, the Beijing-based lab formerly known as Zhipu AI. Z.ai has launched its latest open-weight model, GLM-5.3-Flash, previously code-named Ox Alpha, saying that the system ran entirely on a cluster of 100,000 domestically produced chips during a high-profile stealth trial. The release is a 320B parameter mixture-of-experts model with 18B active parameters, natively multimodal, with a 1M-token context window, published under the MIT License so teams can self-host, fine-tune, and deploy without commercial restrictions. The launch matters because it puts a name to a model that developers had been probing anonymously for over a week. Ox Alpha, the anonymous "stealth" model that showed up on OpenCode and OpenRouter with no branding attached, is GLM-5.3-Flash, the first natively multimodal entry in the GLM-5 line. During that preview period, the lab behind the model was able to serve 100 trillion free tokens per day , and Z.ai is now emphasizing that the whole stack ran on Chinese AI accelerators rather than Nvidia hardware, a point aimed squarely at Chinese enterprises navigating export controls and at global developers weighing price and availability. Z.ai is the international brand for a company with deep academic roots. Founders Tang Jie and Li Juanzi, both professors at Tsinghua University's Department of Computer Science and Technology and active participants in the university's Knowledge Engineering Group (KEG), established Zhipu AI in 2019 at Tsinghua University Science Park in Beijing. For founders and operators building on open models, GLM-5.3-Flash gives them a permissively licensed, long-context multimodal option positioned to compete on cost against closed frontier APIs, backed by a lab that has now demonstrated it can train and serve at frontier scale on domestic silicon.
Comments (7)
320B-A18B under MIT is the actual headline and it's buried in bullet three. Whoever wrote this tweet needs a raise and also a demotion.
Ox Alpha was the codename? You had the coolest preview brand in the game and swapped it for GLM-5.3-Flash. Marketing crime.
Running entirely on Chinese chips is the sentence the whole industry is going to be quoting tomorrow. Curious if you plan to publish throughput numbers or leave that for the reverse engineers.
Every model this year is 'natively multimodal with 1M context' at this point. What does natively actually mean when half the labs use the word like seasoning?
A CFO reading 'highly competitive price' with no numbers attached is like a cat reading a menu. Give us the per-million-token figure or we assume the worst.
Any word on rate limits for the free tier and whether the 1M context is actually usable or one of those 'technically supported' situations that OOMs at 200k?
MIT license on a 320B model is a genuinely wild move and I don't think people have processed it yet. The fine-tune community is about to eat very well.