A model that can read an entire codebase, plan for hours, and ship a working app used to be a closed, cloud-only luxury. A Beijing lab is now handing the weights to anyone — and making sure they run on chips made in China.
Key takeaways
- Zhipu (智谱) open-sourced GLM-5.2 on 17 June 2026 with a stable 1M-token context and the permissive MIT license, publishing weights on Hugging Face and ModelScope.
- GLM-5.3-Flash (320B-A18B), released 26–27 August 2026, is the first native multimodal member of the GLM-5 family and scored 57 on the Artificial Analysis Intelligence Index — level with Anthropic's Claude Opus 4.8 in that ranking.
- Both models were adapted on day zero to domestic accelerators including Huawei Ascend (昇腾), Cambricon (寒武纪) and Moore Threads (摩尔线程); GLM-5.3-Flash was served entirely on roughly 100,000 homegrown chips.
- Zhipu itself became the world's first publicly listed AGI foundation-model company when it listed on the Hong Kong Stock Exchange (02513.HK) on 8 January 2026, raising more than HK$4.3 billion (≈ US$550 million).
- The combination — open weights plus a domestic hardware stack — is a deliberate bet that frontier capability and sovereign compute can be delivered together.
From a chat tool to an "agentic engineer"
Zhipu's story this year is less about one breakthrough and more about a strategy: keep the flagship models open, and make them run where China's own silicon is. GLM-5.2, launched in June, is built around "long-horizon tasks" — the idea that a model should not just answer a prompt but work through a multi-hour engineering job: develop, integrate, test, and package a deliverable. The headline technical claim is a 1M-token context that stays reliable under real load, not just in a specification. In Zhipu's framing, that lets the model hold an entire novel, a 500-page PDF, or a mid-size software repository in memory at once.
On public coding leaderboards the model landed within roughly 1–4% of the strongest closed competitor, Claude Opus 4.8, and Zhipu said it reached first place among available models on Code Arena, a blind-test coding benchmark with over a million users. Those are the company's own cited results, and they should be read as marketing-adjacent until independently reproduced — but the architectural direction is consistent with where the whole field is moving.
The domestic-chip handshake
What makes the GLM-5.2 release notable for readers outside China is the hardware footnote. Zhipu said the model was adapted on day zero to eight domestic accelerator platforms: Huawei Ascend (昇腾), T-Head (平头哥), Moore Threads (摩尔线程), Cambricon (寒武纪), Kunlunxin (昆仑芯), MetaX (沐曦), Hygon (海光) and Biren (壁仞). In other words, the open model was not an afterthought bolted onto Nvidia GPUs; it was designed to run on Chinese chips from the start.
GLM-5.3-Flash pushed this further. Released in late August as the first native multimodal model in the GLM-5 line, it carries 320B parameters with 18B active (320B-A18B) and can take image, video and design inputs and generate code from them. It was served entirely on a cluster the company described as roughly 100,000 domestic chips. Before its official unveiling it was tested anonymously under the nickname "Ox-Alpha" (牛来) on OpenRouter and OpenCode, where it topped both platforms and logged about 62 trillion token calls. Zhipu priced it at roughly a tenth of GLM-5.3 and, during a launch discount, about a fortieth of Opus 4.8 — a price point that only works if the inference cost on homegrown silicon is genuinely low.
Why a listed company keeps giving code away
The open-source pace is striking given Zhipu's other 2026 milestone: on 8 January it listed on the HKEX at HK$116.20 per share, opened at HK$120, and reached a market value around HK$528 billion (≈ US$68 billion). It billed itself as the world's first AGI foundation-model company to go public. The IPO raised more than HK$4.3 billion (≈ US$550 million), with 11 cornerstone investors subscribing HK$2.98 billion.
A newly public company might be expected to guard its best models. Zhipu has done the opposite, releasing its strongest weights under MIT — the most permissive common license, allowing free commercial use. The logic appears to be ecosystem: an open model that runs on domestic chips pulls developers, partners, and eventually paid cloud usage toward Zhipu's own stack. The GLM architecture already adapts to more than 40 domestic chip types, according to the company, which positions open release as a distribution channel rather than a giveaway.
Honest limitations
This piece relies on Zhipu's official announcements, the company's research pages, and Chinese state and financial media (China Daily, Science and Technology Daily, China Securities). Independent third-party benchmark reproductions of the 1M-context reliability and the 1–4% gap to Opus 4.8 are limited. The "100,000 domestic chips" figure comes from press reporting of the company's disclosure rather than a published cluster specification. Financial figures for the IPO are drawn from listing-day coverage and may differ slightly from final filings. I have not tested the models myself, and open-weight availability does not guarantee ease of deployment on every named chip.
What readers can do now
- If you build with language models, try GLM-5.2 or GLM-5.3-Flash via Hugging Face or ModelScope and benchmark them on your own long-context tasks rather than trusting vendor numbers.
- If you deploy in China or on constrained hardware, evaluate the day-zero domestic-chip adapters (Ascend, Cambricon, Moore Threads) as a hedge against single-vendor GPU supply risk.
- Watch whether Zhipu's "open weights + sovereign silicon" model produces durable developer adoption — that, not any single benchmark, will show if the strategy works.