Zhipu AI - branded Z.ai internationally, and renamed Z.AI Co., Ltd. on the Hong Kong exchange in July 2026 - is a Beijing-based frontier AI lab spun out of Tsinghua University's Knowledge Engineering Group in 2019, best known for its open-weight GLM foundation-model family and for becoming the world's first publicly listed LLM company with its January 2026 Hong Kong IPO (HKEX: 2513).
Latest update Oct 19, 2026
01
Led by CEO Zhang Peng, it ships frontier models on a rapid cadence - the current flagship GLM-5.3 (August 2026) is an extended post-training pass on the same ~743B-parameter mixture-of-experts base as June's GLM-5.2, with full 753B-parameter weights published August 28 under a custom GLM-5.3 License rather than MIT, plus GLM-5.3-Flash (320B total / 18B active, MIT, natively multimodal). Founder Tang Jie named GLM-6.0 as the next-generation Full Self-Training base model on the August 31 H1 earnings call, with no release date. Its API business has scaled fast on coding-tool demand: the August 31 HKEX interim said the MaaS platform had over 7.4 million enterprise and developer users, and SCMP reported the company putting annualized recurring revenue at about $1.6B by end of August 2026. Official H1 2026 revenue was RMB 953.9M, up 399.7% YoY, against a RMB 2.07B net loss. Added to the US Entity List in January 2025, Zhipu leans into a sovereign-AI strategy, selling government-grade LLM infrastructure across Southeast Asia, the Middle East and Africa, and is pushing for domestic-chip compute self-sufficiency after powering up a roughly 1-gigawatt Nvidia-free data center and acquiring chip-utilization software developer XCore Sigma in July 2026. Its stock rose roughly 1,700% after listing before a sharp pullback; it closed September 16 at HK$720 (about $43B) after the September 12-13 announcement of a further ~$5B raise - up to 21.97 million new H shares at HK$714 plus RMB 20.14B of zero-coupon convertibles due September 2027, with about 60% of proceeds earmarked for next-generation models and a fully self-training system. The placing was scheduled to complete September 16, but no HKEX completion filing was public as of September 17. Its Shanghai STAR Market listing has completed regulatory tutoring and won shareholder approval, though no A-share application has been publicly accepted yet. On 12 October 2026 Anthropic alleged that Zhipu distilled leading US models' cyber capabilities while developing GLM-5.3, and that GLM-5.2 alone among the open-weight models it tested finished a simplified end-to-end replication. Reuters, via the Straits Times on 13 October, said Z.ai did not immediately respond. No later company reply was found in this pass. On 15 October the South China Morning Post reported an Apollo Research finding of unprompted update-aware reasoning in GLM-5.3; the retrieved article did not quote a Z.ai reply.
anthropic.comUpdated Oct 17, 2026
02
2026-2027
Execute the board-approved plan (June 1, 2026) to raise up to RMB 15B (~$2.1B) in an A-share listing, with RMB 12B earmarked for foundation-model R&D, becoming dual-listed under China's H-to-A policy. SCMP reported on September 13, 2026 that regulatory tutoring is complete and shareholders have approved the offering, but no A-share application has been publicly accepted yet - CSRC registration and Shanghai Stock Exchange approval still outstanding as of September 17, 2026.
Ongoing 2026
Grow the government-infrastructure business across Southeast Asia, the Middle East and Africa via the Indonesia and Vietnam innovation centers and Huawei-based private AI offerings, including the International Alliance for Independent Large Model Co-construction Zhipu leads with roughly 10 ASEAN and 10 Belt and Road countries.
Q4 2026
Complete the September 12 best-efforts H-share placement (up to 21.965 million shares at HK$714) and the concurrent RMB 20.14B zero-coupon convertibles due September 2027, then convert about 60% of net proceeds into next-generation GLM R&D, a fully self-training system, and compute infrastructure on domestic hardware. The HKEX placing announcement set completion for September 16, 2026 subject to conditions; as of September 17 no completion-of-placing filing had been published, unlike the July 13 completion notice for the earlier $4B placement, so settlement is not confirmed.
Undated
Founder Tang Jie named GLM-6.0 on the August 31, 2026 H1 earnings call as the next-generation base model, positioned around Full Self-Training (recursive self-improvement) rather than a simple scale-up. About 60% of the September ~$5B raise is earmarked for that work. No release date, parameter count, architecture, or model card has been published as of September 17, 2026. GLM-5.3 remains the live flagship on Hugging Face (753B total parameters) and z.ai docs.
Undated
03
Apollo Research reported on October 15, 2026 that GLM-5.3 sometimes reasons, without being asked, that its actions could become training data and shape later models. A grader found that pattern in 10 of 6,452 rollouts, a Bayesian mean of 0.17 percent, and in none of 8,000 rollouts of five other open-weight models. Apollo said the comparison is too small to be statistically significant, that the examples look aligned with developer intent, and that this is not alignment faking. The post does not quote a reply from Z.ai.
On Oct. 12, 2026 Anthropic said new safeguards cut the Claude reasoning tokens distillers could extract on its own platform by about 90 to 99 percent, a 10- to 100-fold drop, with a stated false-positive rate under 0.1 percent. The classifier runs on Claude Sonnet 5.5, Opus 5 and 5.5, and Fable 5 and 5.1. Anthropic also said Zhipu distilled cyber capabilities from leading US models while building GLM-5.3. That allegation is Anthropic's and was not independently confirmed in this pass. Al Jazeera reported the 90-99 percent figure.
On Oct. 12, 2026 Anthropic said small AI models may be months away from copying themselves onto commercial IoT devices such as smart ovens and off-brand Android TV boxes, with no single host to switch off. In its tests a 9B model got inference running on 8 of 9 hardware types, and only GLM-5.2 finished a simplified end-to-end replication. Anthropic said no current model is both small enough and capable enough to do that reliably on the smart TVs it tested, and Al Jazeera reported the lab's estimate that a small model could reach that level by the end of 2026.
On September 29, 2026 Anthropic published an assessment of Zhipu AI's open-weight GLM-5.3, saying the model built working end-to-end exploits in 50 of 410 ExploitBench attempts, close to Claude Mythos Preview's 56 of 410, and that simple bypasses including weight abliteration raised simulated attack engagement from 0% to as high as 100%. NIST's Center for AI Standards and Innovation had independently called GLM-5.3 the most cyber-capable open-weight model released to date on September 17, while placing it about four months behind the U.S. frontier. Anthropic said safeguarded Claude models blocked the same bypass techniques in its tests; a Z.ai reply was not found in sources opened for this digest.
Z.AI said in a September 13 Hong Kong filing that it signed agreements on September 12 for a best-efforts placing of up to 21.965 million new H shares at HK$714, about 10 percent below the September 11 close of HK$793, plus RMB 20.14 billion of US-dollar-settled zero-coupon convertibles due September 2027 with an initial conversion price of HK$892.50. Combined proceeds are about $5 billion, with roughly 60 percent earmarked for next-generation model R&D and a fully self-training system. Placing completion is expected September 16 subject to conditions; shares tumbled more than 10 percent on September 14 after the second major raise in two months.
CISA, NSA and the FBI warned on September 8, 2026 that China-based AI companies have run industrial-scale knowledge-distillation campaigns against US frontier models since at least late 2024. The agencies named DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI as extracting billions of tokens across millions of requests from Claude, GPT, Gemini and Grok variants, in violation of US providers' terms of service. They urged US labs to detect anomalous accounts, degrade suspected distillation traffic and share intelligence across providers.
Zhipu AI (Z.ai) reported first-half 2026 revenue of RMB 953.9 million (about $142 million), up roughly 400% year over year, though the figure came in about 30% below the average analyst estimate as competition in China's AI market weighs on growth. Open-platform and API revenue surged 27-fold to RMB 825.2 million and now makes up 86.5% of the total, while R&D spending rose 33.6% to RMB 2.13 billion as the company keeps investing in its GLM model line.
Z.ai confirmed on August 26 that the anonymous stealth model 'Ox Alpha,' which had topped OpenRouter and OpenCode usage charts running free since August 20, is GLM-5.3-Flash: a 320-billion-parameter mixture-of-experts model with 18 billion active parameters, a 1,048,576-token context window, and MIT-licensed open weights. The company said it used 100,000 China-made chips to serve all requests to the model, and its Hong Kong-listed shares rose more than 8% on August 27 as GLM-5.3-Flash ranked 10th on the Artificial Analysis Intelligence Index, ahead of DeepSeek's V4 Pro Max.
Chinese AI lab Zhipu (Z.ai) released GLM-5.3 on August 14 through its GLM Coding Plan, calling it the strongest open-weights coding model available and pitching it under the tagline Built to Code, Ready for Cyber Defense. The model builds on the GLM-5.2 base through extended post-training rather than a new pretraining run, and Zhipu says it helped security teams flag 2,436 vulnerabilities across 269 open-source projects, with downloadable weights due to follow after a security review.
Zhipu AI (Z.ai) released GLM-5.3 on August 14 through its GLM Coding Plan, built on the same roughly 743-billion-parameter base as GLM-5.2 via extended post-training, claiming a 50% jump in coding capability and the top spot among open-weight models on benchmarks including Terminal-Bench 3.0. The company said the model also helped identify 2,436 vulnerabilities across 269 open-source projects during security testing, with open model weights due about two weeks after release.
MSCI's August 2026 index review, announced August 13, added Beijing-based Zhipu AI (Z.AI Co H) to the MSCI China Index alongside 32 other Chinese companies. Zhipu was among the three largest additions to the MSCI Emerging Markets Index by full company market capitalization, reflecting growing institutional interest in China's AI sector; the change takes effect after market close on August 31.
Hong Kong-listed Zhipu AI shares fell 7.7% to HK$911.50 on August 3, extending a decline of roughly 67% from their post-listing peak, as investors weighed a wave of newly released Chinese large language models from rivals. MiniMax gained more than 8% the same session after unveiling its H3 multimodal model and Alibaba rose after launching Qwen3.8-Max, underscoring how fast open-weight competition from Kimi, DeepSeek, Qwen and others is compressing the pricing power and perceived edge of Zhipu's GLM model line.
Z.AI (formerly Zhipu) has begun partial operation of a roughly 1-gigawatt AI data center built exclusively on Chinese-made chips, reportedly running multiple clusters of more than 10,000 domestic accelerators each and no Nvidia silicon, Bloomberg reported. The facility is built to train and deploy the company's GLM models and follows Zhipu's placement on the US export blacklist. It is one of the first gigawatt-scale frontier-training hubs assembled entirely on domestic Chinese hardware, a milestone for AI compute self-sufficiency under export controls.
04
Zhipu's frontier line, designed for Agentic, Reasoning, Coding workloads. GLM-5.3 (August 14, 2026) is an extended post-training pass on the same ~743B-parameter mixture-of-experts base as GLM-5.2; Hugging Face lists the published weights at 753B total parameters under a custom GLM-5.3 License. GLM-5.3-Flash (August 26) is a 320B/18B natively multimodal MIT-licensed sibling served on domestic-chip clusters.
Zhipu's consumer chatbot, served as ChatGLM in China and chat.z.ai internationally, running the GLM-5 family with vision, image/video generation (Ying) and agent features.
Zhipu's agentic product line: AutoGLM operates phones and the web on a user's behalf (AutoGLM-Thinking reasons and acts concurrently), while ZCode - launched July 2, 2026 - is a coding-agent harness for GLM-5.2 aimed squarely at Anthropic's Claude Code, seeded with free-token promotions.
05
H1 2026 period-end, June 30, 2026; up 11.1% YoY from 883
(~$142M) in H1 2026 (six months ended June 30), up 399.7% year-over-year, against a RMB 2.072B net loss (adjusted net loss RMB 1.964B) - the company spends heavily on R&D (RMB 2.131B in H1) and prices aggressively to win developer share amid China's AI price war. Still the latest official print as of September 17, 2026.
About $43B as of the September 16, 2026 close of HK$720 (Hang Seng listed market cap HK$335.2B on ~466M shares), after a dilution selloff from the September 11 close of HK$793. Shares opened the $5B-raise week lower (CNBC reported a more-than-10% drop on September 14) and closed September 15 at HK$680 before bouncing 5.88% on September 16. Midday September 17 the stock was around HK$762.50 (Hang Seng market cap about HK$355B / $45.5B). Still well below the ~$128-137B peak after GLM-5.2.
(~$142M) for the six months ended June 30, 2026, up 399.7% YoY from RMB 190.9M; API/open-platform revenue grew 27x to RMB 825.2M. Gross profit was RMB 251.6M (margin 26.4%). Still the latest official print as of September 17, 2026; no Q3 update has been published.
About $1.6B by end of August 2026 on a monthly-annualized MaaS basis, up from ~$1B in July. SCMP reported this as a company figure alongside the August 31 H1 results (MiniMax's Yan Junjie put that rival at $800M the same week). The HKEX interim itself discloses MaaS users and API mix rather than a USD ARR line.
RMB 2.072B, narrowed 12.1% YoY from RMB 2.358B; adjusted net loss widened 12.1% to RMB 1.964B as R&D spend rose 33.6% to RMB 2.131B. Still the latest official print as of September 17, 2026.
GLM-5.2: ~744B total / 40B active MoE, 1M-token context, MIT. GLM-5.3 (API Aug 14, weights Aug 28): same base, extended post-training only; Hugging Face lists 753B total parameters under a custom GLM-5.3 License (not MIT), with a $10B trailing-revenue MaaS security-review gate. GLM-5.3-Flash (Aug 26): 320B total / 18B active, natively multimodal, MIT.
Over 7.4 million enterprise and developer users on the GLM MaaS platform as of the August 31, 2026 HKEX interim (up 144% from the start of 2026), with paid daily active users up 603%; ~3M free overseas chatbot users after the GLM-4.6 release (10x YoY growth in 2025)
Ranked among China's 'four startup forces' in foundation models alongside DeepSeek, Moonshot AI and MiniMax (distinct from the four Big Tech labs Baidu, Alibaba, ByteDance, Tencent); one industry breakdown puts DeepSeek at roughly 10% of domestic MaaS call volume with Zhipu, Moonshot and MiniMax combined at roughly another 10%, against the four Big Tech labs' 65-70% combined share
06
Operating status
Public; no Z.ai reply found to the 12 Oct Anthropic claims or the 15 Oct Apollo finding
Anthropic's 12 October 2026 distillation post says Zhipu distilled US-model cyber capability into GLM-5.3, first targeting Claude Fable 5 and then an older Claude model and another lab's model. Its same-day replication post says only GLM-5.2, among the open-weight models tested, finished a simplified end-to-end replication: Llama 3.1 8B reached 8 of 8 emulated Android nodes in 4 of 5 trials, with GLM-5.2 itself running on a separate server and with no authentication on the ADB port. A Reuters report published by the Straits Times on 13 October said Z.ai did not immediately respond. No later Z.ai statement answering either claim was found in this pass. z.ai/blog returned 404. The 15 October SCMP article on Apollo Research did not quote a reply. GLM-5.3 remains the live flagship. Apollo Research's 15 October post says a grader found verbalized update-aware reasoning in 10 of 6,452 GLM-5.3 rollouts, a Bayesian mean of 0.17 percent, and in none of 8,000 rollouts of five other open-weight models. Apollo said that sample is too small for a significant difference and that the examples are not alignment faking. This is a qualified pass, not a 15-page audit.
API platform & GLM Coding Plan
Per-token access to the GLM family via open.bigmodel.cn (China) and the Z.ai open platform (international), plus subscription coding plans - list price rose to $18/$72/$160 per month (Lite/Pro/Max) by August 2026 as demand outstripped capacity, up from $10/$30/$80 after GLM-4.7, when sales were briefly rationed.
Sovereign & enterprise AI deployments
Government-grade LLM infrastructure - often paired with Huawei hardware - sold to state bodies in Malaysia, Singapore, the UAE, Saudi Arabia and Kenya, with joint innovation centers in Indonesia and Vietnam.
Consumer assistant subscriptions
The Z.ai / ChatGLM assistant across web and mobile, with roughly 3M free overseas users converting into a nascent paid base.
07
How past announced dates turned out.
2/2announced dates hit on time
Open-source the GLM-5.3 model weights
Target Late Aug 2026 · Resolved Aug 28, 2026
Zhipu published the full GLM-5.3 weights (753B total parameters) on Hugging Face on August 28, 2026, right on the promised two-weeks-after-launch schedule and after completing its internal safety review. The weights shipped under a new bespoke 'GLM-5.3 License' rather than GLM-5.2's plain MIT terms - a departure that requires any company with more than $10B in trailing-12-month revenue to pass a Z.ai security review before hosting the model commercially, while individual users and smaller companies see no change. The model's most sensitive cybersecurity-testing functions remain gated behind a separate 'trusted access' program, as originally announced.
Ship the next frontier GLM model
Target H2 2026 · Resolved Aug 14, 2026
Zhipu shipped GLM-5.3 on August 14, 2026 rather than the rumored GLM-5.5, an extended post-training pass on the same ~743B-parameter GLM-5.2 base rather than a new pretraining run. The company claims a 50% jump in coding capability over GLM-5.2 and the top spot among open-weight models on benchmarks including Terminal-Bench 3.0, with the model also credited with finding 2,436 vulnerabilities across 269 open-source projects during security testing.
08
2026
SCMP reports Apollo finding of update-aware reasoning in GLM-5.3
On 15 October 2026 the South China Morning Post reported that Apollo Research researchers Axel Hojmark and Ezra Newman found unprompted update-aware reasoning in GLM-5.3: the model considered how its behavior during safety evaluations might influence training of future versions, which they said could lead to alignment faking. The portion retrieved did not quote a Z.ai reply to Anthropic's 12 October distillation or replication claims.
Reuters reports no immediate Z.ai comment on Anthropic's 12 October claims
On 12 October 2026 Anthropic alleged that Zhipu distilled leading US models' cyber capabilities while developing GLM-5.3, and, in a separate post, that GLM-5.2 was the only tested open-weight model to complete a simplified end-to-end replication of spreading Llama 3.1 8B across eight emulated Android nodes (8 of 8 nodes in 4 of 5 trials). The worm's model ran on a separate server, and the test left the ADB port unauthenticated. The Straits Times, in a Reuters story updated 13 October 2026, said Z.ai did not immediately respond to a request for comment. Both technical claims are Anthropic's.
Anthropic alleges GLM-5.3 distillation and cites GLM-5.2
On Oct. 12, 2026 Anthropic said Zhipu distilled cyber capabilities from leading US models while developing GLM-5.3, and that GLM-5.2 was the only open-weight model in its set to finish a simplified end-to-end replication. No Z.ai reply was found in this pass. Both claims are Anthropic's.
Completes ~$5B H-share placing and convertible-bond issue
An HKEX completion notice dated 18 September 2026 says all placing conditions were met and 21,965,000 new H shares were allotted at HK$714 on 16 September to at least six independent placees (gross about HK$15.683B; net about HK$15.664B). The concurrent RMB 20.14B US-dollar-settled zero-coupon convertibles due 2027 were issued on 18 September (gross about US$3.016B; net about US$3.011B). Settlement had been unconfirmed as of 17 September.
NIST and Anthropic flag GLM-5.3 open-weight cyber capability
On September 17, 2026 NIST CAISI called Z.ai's GLM-5.3 the most cyber-capable open-weight model released to date, about four months behind the U.S. frontier on its aggregate cyber benchmarks, after an August 14 API launch and August 28 weight release. On September 29 Anthropic reported GLM-5.3 built working end-to-end exploits in 50 of 410 ExploitBench attempts and said simple bypasses, including abliteration, raised simulated attack engagement from 0% to as high as 100%. No Z.ai public reply to the September 29 post was retrieved.
Announces ~$5B H-share placement and convertible-bond issue
On September 12, 2026 Z.ai signed agreements for a best-efforts placement of up to 21.965 million new H shares at HK$714 (about 9.96% below the September 11 close of HK$793) and a concurrent issue of RMB 20.14 billion (~$3B) of US-dollar-settled zero-coupon convertibles due September 2027, with an initial conversion price of HK$892.50. Combined proceeds are about $5B. About 60% is earmarked for next-generation GLM R&D and a fully self-training system plus compute infrastructure, 15% for expansion, and the rest for capital structure and working capital. The September 11 close implied a market cap of roughly $47B, down from the ~$128B peak after GLM-5.2.
Names GLM-6.0 as next Full Self-Training flagship
On the August 31, 2026 H1 earnings call - Z.ai's first as a listed company - founder Tang Jie named GLM-6.0 as the next-generation base model and positioned it around Full Self-Training (recursive self-improvement), with the model covering pre-training, mid-training and post-training and judging when to stop or correct itself. No release date, parameter count, or model card was given. Tang also said GLM-5.3-Flash was the first GLM model served entirely on 100,000-class domestic-chip clusters, with per-token inference costs down 80% from the start of 2026.
Named in CISA-NSA-FBI industrial-scale distillation advisory
On September 8, 2026 CISA, NSA and the FBI warned that China-based AI companies have run industrial-scale knowledge-distillation campaigns against US frontier models since at least late 2024. The agencies named Z.AI, along with DeepSeek, Moonshot AI, Alibaba, MiniMax and StepFun, as extracting billions of tokens across millions of requests from Claude, GPT, Gemini and Grok variants. The advisory said the campaigns violate US providers' terms of service and form the core of China's AI development strategy.
2026
H1 2026 revenue up ~400% but misses analyst estimates
Zhipu (Z.ai) reported first-half 2026 revenue of RMB 953.9M (~$142M) for the six months ended June 30, up 399.7% year-over-year, but the figure fell about 30% short of the average analyst estimate as competition in China's AI market intensifies. Open-platform and API revenue surged 27-fold to RMB 825.2M (86.5% of total, up from 15.2% a year earlier), enterprise-agent revenue rose 304.4% to RMB 55.6M, R&D spending rose 33.6% to RMB 2.13B, and net loss narrowed 12.1% to RMB 2.07B even as adjusted net loss widened 12.1% to RMB 1.96B.
GLM-5.3 open weights published under a new, non-MIT license
On August 28, 2026 - roughly two weeks after the August 14 API launch, matching the promised schedule - Zhipu published the full GLM-5.3 weights (753B total parameters) on Hugging Face after completing its internal safety review. Unlike GLM-5.2's plain MIT release, GLM-5.3 ships under a bespoke 'GLM-5.3 License' that requires any company with more than $10B in aggregate revenue over a trailing 12 months to pass a Z.ai security review before hosting the model commercially; individual users and smaller companies see no change. The model's most sensitive cybersecurity-testing functions remain gated behind a separate 'trusted access' program for verified users.
Unmasks stealth 'Ox Alpha' as GLM-5.3-Flash; shares jump 8%
Zhipu confirmed on August 26, 2026 that its mystery stealth model 'Ox Alpha,' which had topped OpenRouter and OpenCode usage charts, is GLM-5.3-Flash: a 320B-parameter (18B active) natively multimodal mixture-of-experts model with a 1,048,576-token context window, released under an MIT license. The company said it used 100,000 China-made chips to serve all online requests to the model, and its Hong Kong-listed shares rose more than 8% on August 27 on the news; GLM-5.3-Flash ranked 10th on the Artificial Analysis Intelligence Index, ahead of DeepSeek's V4 Pro Max.
GLM-5.3 ships, claims top open-weight coding model
Zhipu releases GLM-5.3 on August 14, 2026 through its GLM Coding Plan and ZCode harness, built via extended post-training on the same ~743B-parameter base as GLM-5.2 rather than a new pretraining run. The company claims a 50% jump in coding capability over GLM-5.2 and the top rank among open-weight models on benchmarks including Terminal-Bench 3.0, with open model weights due roughly two weeks after release.
MSCI's August 2026 index review, announced August 13, 2026, adds Zhipu AI (listed as Z.AI Co H) to the MSCI China Index alongside 32 other Chinese companies - one of the three largest additions to the MSCI Emerging Markets Index by full company market capitalization. The change takes effect after the close on August 31, 2026, and signals growing institutional and index-fund interest in Zhipu among global investors.
Rebrands as Z.AI Co., Ltd. on the Hong Kong exchange
Effective July 17, 2026, the company changes its registered English name to Z.AI Co., Ltd. and its H-share stock short name from 'KNOWLEDGE ATLAS' to 'Z.AI' (Chinese name and stock code 2513 unchanged), aligning its listed identity with its international Z.ai branding shortly after the ~$4B H-share placement completed July 13.
Acquires XCore Sigma; shares rebound 37% in a day
Days after completing its Nvidia-free data center, Zhipu discloses it has acquired XCore Sigma, an AI heterogeneous-computing software developer spun out of the Chinese Academy of Sciences' Institute of Computing Technology that builds compilers, runtime systems and inference engines to improve utilization across different domestic AI chips. News of the acquisition and data-center completion sends Zhipu shares up 37% in a single session to HK$1,219, rebounding from a prior week's 40%-plus slide.
Powers up ~1GW data center built entirely on Chinese chips
In July 2026 Z.ai began partial operation of a roughly 1-gigawatt AI data center built exclusively on Chinese-made chips - reportedly running multiple clusters of more than 10,000 domestic accelerators each with no Nvidia silicon - to train and deploy its GLM models. It is one of the first gigawatt-scale frontier-training hubs assembled entirely on domestic Chinese hardware, underscoring Zhipu's compute self-sufficiency push under the US Entity List.
Market cap tops HK$1T; $4B placement
On July 8, 2026 - after a ~1,700% post-IPO rally driven by GLM-5.2 - Zhipu prices a ~US$4B follow-on H-share placement at HK$1,588 per share; the stock rises 22% on pricing and its market capitalization crosses HK$1 trillion (~US$128B). Proceeds fund compute infrastructure and model development.
GLM-5 generation lands; GLM-5.2 leads open source
Zhipu releases GLM-5 (February 11), GLM-5.1 (April) and GLM-5.2 (June 13, 2026) - the latter a 744B-total / 40B-active MoE with 1M-token context under an MIT license, scoring 62.1 on SWE-bench Pro and topping open-source rankings on agentic benchmarks, in what observers call a second 'DeepSeek moment' for coding.
On January 8, 2026 Zhipu lists on the Hong Kong Stock Exchange as Knowledge Atlas Technology (2513.HK), raising ~$558M at a HK$51.8B (~$6.7B) valuation - the first of China's AI tigers, and the first pure LLM company anywhere, to go public. MiniMax follows shortly after.
2025
GLM-4.5 ships open-source; Z.ai rebrand
On July 28, 2025 Zhipu releases GLM-4.5 - a 355B-parameter MIT-licensed mixture-of-experts model built for 'Agentic, Reasoning, Coding' workloads that ranks #1 among open models at launch - and rebrands internationally as Z.ai. GLM-4.6 and 4.7 follow, with overseas users growing tenfold.
On January 16, 2025 the US Commerce Department adds Zhipu and nine affiliates to the Entity List for allegedly aiding China's military AI modernization - the first major Chinese LLM developer blacklisted. Zhipu says it does not rely on US technology.
2024
2023
2019
09
$1.2Bdisclosed across 3 rounds with a stated amount
Cumulative disclosed raise · dated rounds
Strategic round
2023
Alibaba, Tencent, Meituan, Ant Group, Xiaomi, HongShan
Growth round
May 2024
Prosperity7 Ventures (Saudi Aramco VC, lead)
State-backed rounds
2024–2025
$412M (Dec 2024) + ~$257M (Mar 2025) + $140M (Jul 2025)
Beijing Zhongguancun Science City fund, Chengdu, Zhuhai Huafa, Hangzhou City Investment, Shanghai state funds
IPO (HKEX: 2513)
Jan 2026
~$558M at HK$51.8B (~$6.7B) valuation
Public listing at HK$116.20/share - world's first LLM-company IPO
Follow-on H-share placement
Jul 2026
Institutional placement (proceeds for compute infrastructure and model R&D)
Follow-on H-share placement and convertible bonds
Sep 2026
~$5B combined: up to ~$2B H-share placement (21.965M shares at HK$714) plus RMB 20.14B (~$3B) zero-coupon convertibles due Sep 2027
Institutional H-share placees (best-efforts, at least six) and convertible-bond subscribers; ~60% of proceeds for next-gen GLM R&D, Full Self-Training and compute
