NeuroAI NEUROAINEUROAI.SITE
China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories
ESC

Zhipu becomes the first large-model company to go public — and open-sources GLM-5

Beijing's Zhipu (智谱) listed in Hong Kong as the first large-model stock, then open-sourced its 744B-parameter GLM-5 under the MIT licence.

2026-11-15 · 884 words · NeuroAI
Zhipu becomes the first large-model company to go public — and open-sources GLM-5

A trading-floor screen in Central, Hong Kong, flashed a new stock code at the market open: 02513.HK. The company behind it was not a bank or a manufacturer but a Beijing AI lab that builds large models (大模型). Within five weeks, that same lab would hand its most powerful model to the world for free.

That sequence — ring the bell, then give the weights away — is the story of Zhipu AI (智谱), and it captures a tension running through China's entire frontier-model sector in 2026.

From research lab to listed company

Zhipu listed on the Hong Kong Stock Exchange on January 8, 2026, at an offer price of HK$116.20 (≈ US$14.90) per share. Local financial outlets called it the world's first large-model stock (全球大模型第一股), a claim that hinges on Zhipu being the first company whose core business is building general-purpose large models (大模型) to complete a public listing.

The numbers around the debut were large. The initial public offering raised about HK$4.35 billion (≈ US$558 million), and shares closed their first day at HK$131.50, up 13.2%, valuing the company at roughly HK$57.9 billion (≈ US$7.4 billion). Reporting during 2026 trading noted the stock at one point pushed its market value above HK$1 trillion (≈ US$128 billion) during 2026 trading. Zhipu has also filed to add an A-share listing on Shanghai's STAR Market, a plan that would let it raise around ¥15 billion (≈ US$2.1 billion) and build an "A+H" capital platform.

The listing matters less for the celebration than for what it signals: a model company can now fund itself through public markets rather than only venture rounds.

What GLM-5 actually is

On February 11–12, 2026, Zhipu released and open-sourced GLM-5, the model the company had teased on its listing day. It is an open-weight (开源权重) Mixture-of-Experts system published under the MIT licence on Hugging Face.

  • Total parameters: 744 billion, with 40 billion activated per token.
  • That is roughly double the scale of its predecessor, GLM-4.5 (355B total, 32B active).
  • Pre-training data grew from 23 trillion to 28.5 trillion tokens.
  • Architecture adds DeepSeek Sparse Attention (DSA), multi-token prediction, and a distributed reinforcement-learning framework Zhipu calls "Slime."
  • Context handling extends to a 200K-class window, long enough for whole codebases and long agent (智能体) runs.

Zhipu frames GLM-5 as a tool for "Agentic Engineering" — finishing multi-day software projects, not just writing snippets. Before the official name surfaced, an anonymous build labelled "Pony Alpha" circulated on the OpenRouter platform and drew attention from overseas developers who shipped working apps with it.

The benchmark claims, and the caveats

Zhipu says GLM-5 ranks fourth globally and first among open-weight (开源权重) models on the Artificial Analysis intelligence index — a claim repeated across the company's own materials and independent write-ups. On coding, Zhipu reports 77.8% on SWE-bench-Verified and 56.2 on Terminal Bench 2.0, both described as open-source best scores, with coding quality it says approaches Claude Opus 4.5. In a simulated vending-machine business test (Vending Bench 2), GLM-5 ended a one-year run with a US$4,432 balance, which Zhipu says leads all open models.

Those figures are Zhipu's own and should be read as vendor-reported. Independent leaderboards disagree on exact scores because each uses different scaffolding and attempt counts. The safer read: GLM-5 is a genuine open-weight (开源权重) frontier contender, not a small model with a loud press release.

Why open weights still matter

GLM-5 was also adapted for seven domestic accelerator platforms — Huawei Ascend (昇腾), Moore Threads (摩尔线程), Cambricon (寒武纪), Kunlunxin, MetaX, Enflame, and Hygon — so it can run on Chinese-made hardware rather than only Nvidia GPUs. For buyers with data-residency rules, that combination — permissive licence plus local silicon — is the actual product.

According to Li An, chief scientist at BrainNet (脑机网), the more interesting move is not the benchmark but the licence: a publicly listed company betting that giving weights away strengthens, rather than weakens, its commercial position.

Honest limitations

This article draws on Zhipu's official release notes, the company's model card, and financial reporting from outlets including Securities Times and China Economic Net; it does not independently reproduce the benchmark or efficiency claims. The SWE-bench, Terminal Bench, and Vending Bench numbers are Zhipu's own and have not been re-run by an outside lab in the sources reviewed. The Artificial Analysis ranking reflects that index's methodology at a point in time and shifts as competitors update. The 200K context figure comes from secondary summaries of the model card rather than a direct parameter in the excerpt reviewed. The A-share plan is a filing, not a completed listing, and remains subject to regulatory approval. No claim is made that GLM-5 exceeds closed models on every dimension — its known weaker spots include raw scientific-reasoning scores versus the strongest proprietary systems.

What readers can do now

  • Builders: download the GLM-5 weights from Hugging Face and benchmark them on your own coding and agent (智能体) workloads before trusting vendor scores.
  • Buyers with compliance needs: test GLM-5 on a domestic accelerator stack (Huawei Ascend / 昇腾 or Cambricon / 寒武纪) to see whether local-hardware deployment actually meets your latency and cost targets.
  • Investors: treat the Zhipu listing as a test case for whether "open-weight (开源权重) plus a public balance sheet" is a durable business model, and watch the pending A-share filing for the real capital plan.

Related coverage

Cite this article

NeuroAI (脑机网), "Zhipu becomes the first large-model company to go public — and open-sources GLM-5," 15 November 2026, https://neuroai.site/a/na-model-zhipu-glm5-hk-ipo

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “Foundation Models” → · Back to home · Markdown version

China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories