Chinese Z.AI (Zhipu AI, the company behind the GLM model family) announced on September 13, 2026, the completion of approximately $5 billion in financing. The company will direct about 60% of this amount toward developing the next generation of GLM foundation models and the Fully Self-Training system, in which each new GLM participates in creating the environment, data, and infrastructure for training the next version. The announcement contains no new models, APIs, or benchmarks: this is a capital event, but it directly funds future GLM releases, which should be expected in the coming months.

image

What happened

On September 13, 2026, Z.AI announced the completion of a round of approximately $5 billion. The company raised about $2 billion through the placement of 21.97 million new shares at HK$714 per share, which is about 10% below the previous trading day's close. Another approximately $3 billion was secured through zero-coupon convertible bonds worth 20.14 billion yuan, maturing in September 2027; conversion is possible at HK$892.5, which is 25% above the placement price, and the yield on the instruments is in the range of -0.5% to 0%. According to the company's disclosure, about 60% of the funds will go toward the next generation of GLM foundation models and the Fully Self-Training system, as well as automated generation and filtering of training data, adaptation to domestic chips, and inference optimization. The announcement itself contains no new models or open weights, no benchmarks, no technical reports, and no new APIs.

Context

Z.AI is developing the GLM family of open-weight models, which this year is competing with DeepSeek, and is listed on the Hong Kong Stock Exchange. The round was the second in two months: the company raised about $4 billion in July 2026 and about $5 billion in September, which has cemented the role of the Hong Kong capital market as the main source of financing for Chinese frontier labs that lack access to advanced US chips. The terms of the September deal are themselves telling: zero-coupon convertible bonds with yields of -0.5% to 0% mean that investors are effectively paying for the right to a share in Z.AI's future equity, rather than for interest income. The deal took place against a backdrop of diplomatic tension: two days before its finalization, the NSA, FBI, and CISA in a joint advisory named Z.AI as one of six Chinese companies accused of industrial distillation of American models; Beijing characterized these accusations as unfounded.

Why this matters for the industry

This is a capital event, not a product release, but it clarifies the priorities of one of the leading Chinese labs. The composition of expenses points to a real bottleneck in the industry: in addition to the next generation of GLM and the Fully Self-Training cycle, separate areas are adaptation to domestic chips and inference optimization, meaning the company is investing in computing infrastructure under restrictions on access to American accelerators. The central bet is Fully Self-Training, which the company relates to the direction of Recursive Self-Improvement: if the loop, where the model itself builds the environment and data for the next version, is truly closed, the cost of scaling capabilities will decrease, the frequency of releases will increase, and cheap open-weight models will intensify price pressure on inference overall. For now, the announcement gives teams building products on GLM no direct material to work with: there are no new APIs, weights, or infrastructure releases in it, and the real test of the bet is only possible with the release of the next generation of models and their first public benchmarks.

Why this matters for users

For those already using open-weight GLM models, nothing changes: access, prices, and characteristics of current versions remain the same, and there is no need to change the current production stack based on this news. The practical meaning is different: the next version of the family is funded, and its release and first public benchmarks are likely in the coming months. Expect statements about inference optimization, including throughput, quantization, and token cost, with a caveat about the domestic chip stack. When the new GLM is released, it makes sense to conduct your own A/B test of quality and cost against the current stack and, if necessary, rebuild RAG and agent workflows for a cheaper model. It is also worth remembering the context: the company is under pressure from American regulators, which may affect its access to external markets and services.

What is still unknown / limitations

Fully Self-Training is described as a direction, not a working system: according to TechNode, many engineers are currently involved in this process, and Z.AI's goal is to close the loop as much as possible, meaning the declared cycle from GLM through the environment and data to the next GLM has not yet been closed. The announcement contains no scientific artifacts: no papers, no benchmarks, no model parameters, no description of training methods and evaluation methodology. Details of the distribution of the remaining 40% of funds are not disclosed. The accusations from the NSA, FBI, and CISA remain assertions by US agencies that the company and Beijing dispute, and they have not yet received independent confirmation. Reading the deal structure as a signal of market confidence in Z.AI's technical trajectory is an interpretation, not a fact: from a financial perspective, this is a placement metric. The timing of the next generation of GLM remains an expectation, not an official company commitment.

Sources

Author

Look at AI, editorial team