Zhipu Unveils GLM-6.0 'Fully Self-Training' Plan in HK$39.3 Billion Financing Filing

The filing sets out a new H-share placement and the issue of RMB20.14 billion in zero-coupon convertible bonds. Total net proceeds are expected to be about HK$39.3 billion. About 60%, or HK$23.5 billion, will go to research and development of the next-generation GLM foundation model and the 'fully self-training' system, as well as large-scale training, inference and computing infrastructure deployment and upgrades.

Zhipu defines 'fully self-training' as a closed loop in which the next-generation GLM model is trained in an environment built by the previous-generation GLM. The loop covers three areas: data self-production, environment self-creation and infrastructure self-optimization.

For data self-production, the company plans to use model-to-model self-play, rule verification, execution verification, model review and human spot checks to automatically generate, screen and accumulate high-quality training data, then feed it back into pre-training, mid-training and post-training. For environment self-creation, agents are to collect and convert real-world tasks, try them, generate verifiers and check whether tasks are solvable, expanding the number, type and complexity of training environments so that task environments become scalable and reusable training resources. For infrastructure self-optimization, Zhipu intends to use the model's coding and systems engineering abilities to help develop and optimize operators, kernels, scheduling, caching and the service stack for training and inference, allowing the model to participate in upgrades to the infrastructure that supports its own iteration.

The filing also previews capability targets. Zhipu says it will seek to improve the effective scale of the foundation model and native multimodal unified modeling, deepen effective computational depth without excessively increasing inference costs, and enable longer-chain reasoning, planning and self-checking. Another focus is long-horizon task reinforcement learning. The goal is to build task environments close to real professional work and train models to break down tasks, call tools, interact with environments, recover from errors and verify final results.

On computing, the filing says the market currently offers relatively favorable supply and delivery conditions for high-quality computing resources. Zhipu plans to combine procurement and leasing, invest in chip adaptation, operator development, cluster interconnection and performance tuning, and build a diversified and flexibly allocatable computing supply system. It also plans to invest in a self-developed inference engine and service stack covering multimodal encoding, prompt prefill, per-token decoding, quantization, caching, communication and elastic scaling, with the aim of raising inference throughput and cost efficiency on the same hardware.

Of the remaining proceeds, about 15%, or HK$5.9 billion, is earmarked for business expansion, strategic investment and potential acquisitions. The filing lists selection criteria: targets must be closely related to AI technology and applications, have reliable users or customers and stable operations for no less than two years, and possess a mature product or technology platform with continuous research, development and technology iteration capabilities. The covered directions include large models, intelligent algorithms, computing operations, computing infrastructure, AI platforms, agents and AI-enabled applications. Business expansion will support the MaaS platform, API, Coding Plan and Agent and Co-work products and services. About 25%, or HK$9.8 billion, will be used to optimize the capital structure and supplement working capital. All proceeds are expected to be used by June 30, 2028.

The filing also lays out Zhipu's recent fundraising timeline. On Jan. 8, 2026, it completed a global offering of H shares at HK$116.20 each, issuing about 43.04 million H shares including the over-allotment option and raising about HK$4.896 billion net. By Aug. 31, 2026, those IPO proceeds had been fully spent. On July 9, 2026, Zhipu completed a first placement at HK$1,588 per share, issuing 19.78 million new H shares and raising about HK$31.375 billion net. The placement price was about 13.7 times the IPO price six months earlier. By the end of August, about 34.92%, or HK$10.955 billion, of that placement had been used, leaving about HK$20.42 billion unused. The Sept. 12 financing followed, after the H-share closing price had fallen from its high to HK$793 and the placement price was set at HK$714, down from HK$1,588 in July. Across the three rounds, Zhipu has raised about HK$75.5 billion, or about RMB64.6 billion, net since listing.

The filing says that as business progress accelerated, the pace of using funds under the original fundraising plan also sped up, while computing resources require time from contract signing to deployment and launch. It says the current financing arrangement helps align the pace of new computing deployment with the expansion plan. The convertible bond terms also mention that the conversion price will not be adjusted because of the company's currently expected initial public offering of shares on the Shanghai Stock Exchange STAR Market, a detail that points to a planned A plus H listing.

At an interim results briefing in late August, Zhipu's Tang Jie responded to criticism that the company only does post-training and said the next-generation GLM-6.0 aims at self-evolution. The model and its paper have not yet been released, and the first detailed preview came through the financial filing.