Zhipu spoils GLM-6.0: a fully self-training model, leaked by a filing
The model isn't out and there's no paper yet — but Zhipu's latest funding announcement just put the company's next technical bet in writing, and it's the most explicit commitment to recursive self-improvement any frontier lab has made in a financial disclosure.
Zhipu says roughly 60% of its new HK$3.93 billion raise — about HK$23.5 billion — will fund GLM-6.0 and a "fully self training" system, a term the company defines in the announcement as a closed loop where the model generates its own data, builds its own training environments, and helps optimize the infrastructure it trains on. The three parts are spelled out: data self-generation through model self-play, rule checking, execution verification, model review and human spot-checks, feeding continuously back into pre-, mid- and post-training; environment self-construction, where agents attempt tasks, write their own verifiers, and check whether tasks are solvable, turning task environments into "scalable, reusable training resources"; and infrastructure self-optimization, using the model's coding ability to improve its own operators, kernels, scheduling, caching and serving stack. It is Zhipu's public name for recursive self-improvement — the direction its founder Tang Jie gestured at in late August when he answered critics who say the lab "only does post-training."
The rest of the allocation reads like a war chest: compute purchasing and leasing with chip adaptation and cluster interconnect work, an in-house inference engine covering quantization, caching and elastic scaling, about HK$5.9 billion (15%) for commercial expansion and acquisitions, and about HK$9.8 billion (25%) for the balance sheet — all to be spent by June 30, 2028. Two details from the filing matter as much as the roadmap. First, the funding cadence: this is Zhipu's third equity event since its January IPO at HK$116.20 a share — the July placement priced at HK$1,588, this one at HK$714 — bringing cumulative net proceeds to roughly HK$75.5 billion, with the IPO money already fully spent. Second, the convertible-bond terms contain a clause that the conversion price will not be adjusted for "the expected initial public offering of shares on the Shanghai Stock Exchange STAR Market" — a legal footnote that confirms an A+H listing is being planned. We covered the financing mechanics yesterday — Zhipu raises $5 billion in its second financing in two months.
The honest caveat is in the announcement itself: none of this exists yet. Self-play plus automated verification is exactly the setup critics warn can go self-referential — capability decay or eval distortion once the model is grading its own homework without real-world signal. Zhipu is betting that task environments with executable ground truth (code, tool use, long-horizon professional work) are the external anchor that makes the loop honest. Putting that bet in a HKEX filing, with a dollar figure attached, is a different kind of spoiler than a blog post.
What to watch: whether GLM-6.0 ships with the paper the filing promises, and whether the STAR Market IPO footnote turns into an application before the bonds convert.
Is a self-training loop with no human data ceiling an accelerator or an echo chamber — does it depend entirely on the verifier? Tell us in the comments.
