Z.ai GLM-5.3-Flash Lifts Shares 8% on All-Chinese Chip Serving

What happened

Z.ai, the Beijing lab once known internationally as Zhipu AI, saw its Hong Kong-listed shares rise more than 8% on August 27. The day before, the company had revealed that a mystery model topping developer usage charts was its own GLM-5.3-Flash, and said every online request to that model was being served on 100,000 Chinese-made chips14. On August 26 the company confirmed that "Ox Alpha," an anonymous model that had run for free on OpenRouter and OpenCode since August 20, was GLM-5.3-Flash. The model is a natively multimodal mixture-of-experts system with 320 billion total parameters, 18 billion of them active, and a context window of 1,048,576 tokens. Its weights are released under the MIT license6. At the time it ranked 10th on the Artificial Analysis Intelligence Index, ahead of DeepSeek's V4 Pro Max14.

The short version, "a model running only on Chinese chips," needs some unpacking. CNBC's follow-up coverage and syndicated versions of it describe the claim specifically as serving online requests on domestic hardware1317. That is an inference milestone, and it is a different thing from training. On the company's August 31 earnings call, founder Tang Jie called GLM-5.3-Flash the first GLM model served entirely on 100,000-class domestic-chip clusters. He also said per-token inference costs had fallen 80% since the start of 202614. In practice, the result shows Chinese silicon handling large-scale production traffic for a model developers actually wanted to use. Training on Chinese hardware had already been shown with earlier GLM releases.

Why the chip story moves the stock

The market's reaction makes more sense against Z.ai's history with US export policy. In January 2025 the US Commerce Department put Zhipu and nine affiliates on the Entity List, the first major Chinese LLM developer to be blacklisted, and the company said it did not rely on US technology14. Since then it has built a record of hardware independence one release at a time. In August 2025 it announced that GLM models were compatible with Huawei's Ascend processors, and its GLM-4.6 release that September ran on domestic chips including Cambricon's11. GLM-5, released in February 2026, was described as trained entirely on Huawei Ascend chips without Nvidia GPUs35.

The infrastructure went in before the model. In July 2026 Z.ai began partial operation of a roughly 1-gigawatt data center built only with Chinese-made accelerators. Around the same time it acquired XCore Sigma, a developer of compilers and inference engines meant to improve utilization across different domestic chips, and that news pushed shares up 37% in one session14. Seen this way, the GLM-5.3-Flash announcement was the payoff: proof that the stack could serve a popular model at scale. Investors seem to treat each of these milestones as evidence that sanctions limit Z.ai less than its valuation implies.

Open weights as strategy, with a new asterisk

Open weights are the second half of the story. Z.ai has released its models under the MIT license since July 202511. GLM-5.2, released in June 2026, is a 744-billion-parameter MoE model with 40 billion active parameters, a 1-million-token context window and MIT terms7. One explainer says press coverage reported it was trained on Ascend accelerators with no Nvidia hardware2. Its API price was $1.40 per million input tokens and $4.40 per million output tokens2. The stock reacted far more strongly to that launch than to the August one. Citing the South China Morning Post, one account says shares rose as much as 48% intraday and closed up 32.8% on June 1510.

GLM-5.3 changed the pattern. The flagship, released August 14, is an extended post-training pass on the GLM-5.2 base, not a new pretraining run6. At launch Z.ai offered it only through its Coding Plan and ZCode, saying API access and weights would follow in stages after safety evaluations9. The weights did arrive on August 28, as promised. However, they came under a custom GLM-5.3 License instead of MIT, and that license adds a security review for model-service companies with more than $10 billion in annual revenue36. GLM-5.3-Flash, by contrast, shipped with MIT weights on day one3.

Our reading is that Z.ai is now splitting its catalog in two. Smaller, cheaper models stay fully permissive to win developers. The frontier flagship carries conditions aimed mainly at large commercial rivals. The timing fits that interpretation. On September 8, CISA, the NSA and the FBI named Z.AI among China-based labs they accuse of running industrial-scale distillation campaigns against US frontier models14. A lab under that kind of scrutiny has reasons to keep some control over who profits from its most capable weights.

Where the reporting diverges

The coverage agrees on the main points: the 8% move, the 100,000-chip claim and the open-weight terms. On details, it varies. Release dates differ. GLM-5.2 is dated June 13 in some accounts and June 16 in others72, and GLM-5 appears as February 11, 12 or 13 depending on the source5113. One release tracker lists GLM-5 as proprietary and API-only, while other accounts say it shipped under MIT53. Benchmark claims vary too. GLM-5.2 is cited at 62.1 on SWE-bench Pro in several places, while a local-LLM blog calls it the first open model past 68%78. One explainer's own summary still says GLM-5.3 has no weights, even though its FAQ confirms the August 28 release3. None of this changes the overall story, but readers should treat precise benchmark numbers in this category with caution. GLM-5.2 itself launched without published benchmark scores, so its early performance claims were vendor assertions until later testing10.

On capability, independent rankings put Z.ai near the top of the open-weight field but behind the closed US leaders. One analysis using the Artificial Analysis index puts Anthropic and OpenAI models at 50–53, GLM-5.3 at 44, and Alibaba's Qwen3.8-Max at 4015.

The Gemini contrast

Google's recent releases show how differently the two strategies work. Google shipped Gemini 3.6 Flash in July, 3.7 Flash in August and 3.8 Flash on September 2, a roughly monthly cadence comparable to Z.ai's23. On September 30 it announced Gemini 4 Argon in limited preview. The model raises the output limit from 64,000 to 1,000,000 tokens23. CNET noted that, after months of delays, ordinary users could not yet access the model22. One analysis reports Argon's introductory pricing at $2 per million input tokens and $10 per million output tokens, rising later to $4 and $2028. Both companies are drawing lines around cybersecurity. Google offers a more permissive Gemini 3.8 Flash Cyber through a gated program23, and Z.ai markets GLM-5.3 as "ready for cyber defense" while staging its access9.

The key difference is distribution. Google's frontier models are API-only, and its developer documentation points to Gemma 4 as the open alternative for some retiring workloads26. Z.ai puts frontier-scale weights on Hugging Face. That gives it reach Google's closed models don't have, but it also limits how much of that reach turns into revenue.

The money question

Revenue is the weak point. Z.ai's H1 2026 revenue was RMB 953.9 million, up nearly 400% year over year, against a RMB 2.07 billion net loss14. The 8% gain also didn't hold. In September the company announced a roughly $5 billion raise through a share placement and convertible bonds, its second major fundraising in two months, and the stock fell more than 10%1319. About 60% of the proceeds will go to next-generation models and a "fully self-training system"19. Tang has named GLM-6.0 as that system's flagship but has not given a date14. A Macquarie analyst has warned that Z.ai could stay unprofitable into the next decade15, and Caixin presents the raise as part of a wider Chinese race to stockpile compute18.

The August rally showed investors will reward real independence from Nvidia and strong developer adoption. The September sell-off showed they are also counting what that independence costs. Building a full domestic stack is clearly possible. The open question is whether open weights, which Z.ai mostly gives away, can ever pay for it.