Zhipu Qingyan
Zhipu Qingyan (also known as ChatGLM) is the consumer AI chatbot from Z.ai (formerly Zhipu AI), a Tsinghua University spinoff that became the first major Chinese LLM developer to go public when it IPO'd on the Hong Kong Stock Exchange in January 2026. Powered by the GLM model series (including GLM-5 with 744 billion parameters), it offers AI-powered dialogue, content creation, code generation, and agent capabilities. Available as a standalone app and website, alongside a separate developer API platform; the app reportedly exceeded 25 million users by the end of 2024. China-focused but expanding globally, particularly with developers.
Score generated by AI agents based on publicly cited evidence and reviewed by the project maintainer. Not independently validated. Last assessed 2026-10-24. Score revised 2026-10-24: 38 → 39.
Score History
Timeline events are AI-curated from public reporting. Score trajectory is derived from documented events.
Tang Jie and Li Juanzi founded Zhipu AI in June 2019 as a spinoff of Tsinghua University's Knowledge Engineering Group, focused on knowledge graphs and academic research. Tang's team also helped build the WuDao 2.0 model unveiled in 2021. With no consumer product and little commercial activity, almost none of the enshittification forces applied.
The August 2022 release of GLM-130B marked Zhipu's turn from knowledge graphs to large language models. ChatGLM-130B went online at chatglm.cn in March 2023 alongside the open ChatGLM-6B, whose weights became free for registered commercial use in July 2023. Venture funding grew and China's generative-AI rules took shape, but Zhipu still operated as an open, research-led model developer.
After clearing China's first batch of generative-AI approvals, Zhipu launched Qingyan on August 31, 2023, bringing state political-content rules into a public consumer product. The app improved through 2024 with GLM-4 agent tools, video generation and video calls, passing 25 million users, while Zhipu cut API prices in the 2024 price war. Large rounds from Alibaba, Tencent, Prosperity7 and others funded R&D far beyond revenue.
The U.S. Entity List designation in January 2025 made Zhipu the first Chinese LLM company sanctioned. Growth brought new problems: regulators flagged Qingyan for excessive data collection in May, GLM-4.5 refused protest-related queries in August, and former employees disputed the scale of an autumn restructuring. State-backed funding, the Z.ai rebrand and MIT-licensed releases set up the January 2026 listing.
Z.ai's January 8, 2026 Hong Kong IPO made it the first major Chinese LLM developer to list, and the period since has been defined by demand outrunning compute and repeated price increases. Coding Plan sign-ups were cut by 80% in January, GLM-5 brought price rises and then an apology with refunds in February, API prices rose again with GLM-5.1 in April, and a points-based plan raised new-user fees by up to 261% in July. Capital raising scaled up with follow-on deals in July and September.
Alternatives
Open-source Chinese AI chatbot with reasoning and coding capabilities. Its consumer app is free to use with no subscription required. Easy switch — download the app or use the web interface.
ByteDance's AI chatbot with general-assistant and multimodal features. It surpassed 100 million daily active users during the February 2026 Lunar New Year holiday, according to survey data reported by Reuters. Available as a standalone app, with free access and optional paid features.
General-purpose AI chatbot with text, voice, image and file-analysis features. Mainland China is outside its supported regions. Free tier available; Plus subscription at $20/month.
Dimensional Breakdown
Summaries below were written by AI agents based on the cited evidence. They are editorial interpretations, not independent research findings.
Dimension History
Timeline (53 events)
Zhipu AI founded as Tsinghua University spinoff
Professors Tang Jie and Li Juanzi from Tsinghua University's Department of Computer Science founded Zhipu AI in Beijing's Tsinghua Science Park, initially focusing on knowledge graphs and academic AI research. The company grew directly out of Tsinghua's Knowledge Engineering Group (KEG).
Zhipu co-founder helps unveil WuDao 2.0 trillion-parameter model
Tang Jie, Zhipu co-founder and BAAI researcher, helped develop WuDao 2.0, unveiled in June 2021 with 1.75 trillion parameters and reported as the world's largest AI model at the time. Tang's team had already developed GLM-10B before this unveiling; the WuDao work formed part of the broader Tsinghua and BAAI research ecosystem behind Zhipu's large-model development.
China's algorithm recommendation regulations take effect
The CAC's Internet Information Service Algorithmic Recommendation Management Provisions took effect requiring transparency over recommendation algorithms, user opt-out controls, and algorithm filing. Providers with public opinion properties or social mobilization capabilities faced security assessments and mandatory filing. The provisions established part of the regulatory framework governing Zhipu's subsequent public-facing AI services. In July 2023, Carnegie described China's emerging AI rules, including the recommendation provisions, as among the world's earliest and most detailed.
GLM-130B bilingual model open-sourced
Zhipu and Tsinghua KEG made GLM-130B, a 130-billion-parameter bilingual (Chinese/English) pre-trained model, publicly available with model checkpoints and inference code. The launch page reported performance above GPT-3 175B on LAMBADA benchmarks. The project's October 2022 paper also documented the public release of code, training logs, related tools and lessons learned.
Series B funding from Legend Capital and Qiming Ventures
Zhipu AI announced hundreds of millions of yuan in Series B financing, co-led by Qiming Venture Partners. Legend Capital also co-led the round; the company's investors included Tsinghua Holdings. Qiming said the proceeds would support continued development of high-performance large models with hundreds of billions of parameters.
China adopts Deep Synthesis Provisions for AI content
The CAC, MIIT, and Ministry of Public Security jointly adopted regulations on deep synthesis technology requiring AI service providers to establish algorithm review, ethics review, content monitoring, and personal information protection mechanisms. Service providers with 'public opinion attributes or social mobilization capacity' faced mandatory algorithm registration and security assessments. Taking effect January 10, 2023, these provisions further tightened the regulatory framework constraining Zhipu's subsequent consumer AI services.
ChatGLM-130B and ChatGLM-6B launched simultaneously
Zhipu put ChatGLM-130B online at chatglm.cn and simultaneously open-sourced ChatGLM-6B, a 6.2-billion-parameter model enabling local deployment on consumer GPUs via INT4 quantization. ChatGLM-6B was pre-trained on approximately one trillion Chinese and English tokens, in a 1:1 language ratio according to its launch blog. The team later reported over 10 million Hugging Face downloads across its open-model series in 2023. Economic Observer reported on December 11, 2023 that downloads of the ChatGLM-6B series had exceeded 10 million across open-source communities.
China enacts Interim Measures for Generative AI Services
The CAC, jointly with six other Chinese authorities, announced the Interim Measures for Generative AI Services, effective August 15, 2023. The binding rules require public-facing generative AI services in mainland China to uphold core socialist values and prohibit content such as incitement to subversion or separatism. Providers with public-opinion properties or social-mobilization capacity must complete security assessments and algorithm filing; purely internal enterprise and research use is excluded. Legal scholarship describes the measures as the world's first binding instrument specifically targeting generative AI.
ChatGLM-6B weights opened for free commercial use after registration
Zhipu AI and Tsinghua KEG made the ChatGLM-6B and ChatGLM2-6B weights fully open for academic research and allowed free commercial use once a business registered and received authorization.
Zhipu among first batch approved for public AI services
Zhipu AI was among 11 companies in the first batch approved to offer generative AI services to the public. Under the regulatory framework, public-facing services with public-opinion properties or social-mobilization capacity had to complete security assessments and algorithm filing. Zhipu launched Qingyan at the end of August; its public service was subject to the Interim Measures' political-content requirements.
Zhipu Qingyan consumer AI assistant launched
Zhipu launched Qingyan, its consumer-facing AI assistant based on ChatGLM2, with general Q&A, long-form dialogue, creative writing and code generation. Available as a standalone app and WeChat plug-in, it entered China's growing consumer chatbot market. By December 2024, Zhipu reported that the app had over 25 million users.
Zhipu acquires 94.98% of Beijing Lingxin Intelligent
In September 2023 Zhipu signed transfer agreements to buy an aggregate 94.98% of Beijing Lingxin Intelligent, an AI-services company in which it already held about 5.02%, for RMB73,004,600 in cash. The prospectus says the consideration was fully settled on October 24, 2024, leaving Zhipu with 100% ownership.
Zhipu reports roughly $340 million raised in 2023
Zhipu AI said it had raised more than 2.5 billion yuan (roughly $340 million) in total financing so far in 2023. Investors included Alibaba, Tencent, Meituan, Ant Group, Xiaomi and HongShan, alongside other venture investors and a state-backed fund. The financing made Zhipu one of China's prominent contenders in the large-model market.
GLM-4 All Tools launched with autonomous agent capabilities
Zhipu released GLM-4 with a 128K-token context window, multimodal understanding and the All Tools agent feature, which integrates a web browser, code interpreter and image generation. Subsequent company evaluations reported performance close to GPT-4-Turbo on instruction following and comparable long-context performance to GPT-4-Turbo and Claude 3.
Zhipu workforce reported above 800, with most staff in R&D
A May 2024 Turing Post profile reported more than 800 employees at Zhipu, with roughly 60–70% working in R&D. The profile described the company's ethos as academic and technologically idealistic. Later in 2024, the Center for Data Innovation described Zhipu as China's largest AI startup by headcount and cited a September valuation of approximately $2.8 billion.
Zhipu cuts GLM-3-Turbo API pricing by 80%
Following DeepSeek's May 6 pricing announcement, Zhipu reduced GLM-3-Turbo's API rate by 80%, from 5 yuan to 1 yuan per million tokens, and increased new-user free allocations from 5 million to 25 million tokens. On June 5, Zhipu separately announced GLM-series pricing as low as 0.1 yuan per million-token prompt, amid an industry-wide price war.
Prosperity7 joins reported $400 million Zhipu funding round
Saudi Aramco's Prosperity7 Ventures joined a roughly $400 million financing round for Zhipu AI, valuing it at about $3 billion. The investment made Prosperity7 the first known foreign firm to back a major Chinese generative AI startup, highlighting the geopolitical significance of financing China's large-model developers.
GLM-4 open-sourced with multimodal variant
Zhipu released GLM-4-9B and the multimodal GLM-4V-9B, with multilingual and long-context support. The repository reported GLM-4V-9B performance above GPT-4-Turbo in multiple multimodal evaluations. Model weights follow a custom Model License, while repository code is Apache2.0. By December 2024, the company reported over 30 million global downloads across more than 50 open models, including ChatGLM.
Qingying video generator launched in Qingyan
Zhipu launched Qingying, an AI video generator, on July 26, 2024. It generated videos up to six seconds long at 1440x960 resolution from text or image prompts, according to the company statement cited by Caixin.
Qingyan opens live AI video calls
Alongside the GLM-4-Plus base model, Qingyan's video-call feature went live on August 30, 2024, first for some users with external applications open. It was reported as the first video-call feature from a Chinese large-model company opened to individual users.
Zhipu raises 3 billion yuan in its fifth round of 2024
Zhipu said it had raised 3 billion yuan (US$412 million) in a new funding round, adding to four investment rounds earlier in 2024. Its backers included state-backed investment vehicles, Alibaba, Tencent, Qiming Ventures and Legend Capital.
Zhipu AI and nine affiliated entities added to the US Entity List
The US Commerce Department added Beijing Zhipu Huazhang Technology (Zhipu AI) and nine related entities, including Beijing Lingxin Intelligent Technology, to the Entity List. The final rule, effective January 16, said the entities advanced China's military modernization through the development and integration of advanced AI research. Exports of items subject to the EAR required a license, with a presumption of denial. The January rule did not assign Zhipu a Footnote 4 designation.
Pre-IPO filing discloses Zhipu's joint control structure
Zhipu's pre-listing tutoring documents stated that it had no controlling shareholder. Founder Tang Jie, Chairman Liu Debing and entities acting in concert held almost 37% of voting rights and were identified as the actual controllers.
Qingyan flagged for unauthorized data collection by Chinese regulators
China's National Cyber Security Information Centre flagged Zhipu Qingyan among 35 apps for collecting information beyond what users authorized, following inspections from April 16 to May 15. SCMP dated the notice to May 20; PANews reported the app count and detection period on May 21. The Qingyan finding concerned collection beyond authorization, while the separate finding against Kimi concerned data unrelated to business functions.
COO Zhang Fan confirms departure to start an AI agent company
On June 12, 2025, Zhipu COO Zhang Fan confirmed reports that he would leave to start an AI agent company, which he said was at an early stage and in which Zhipu would invest.
Shanghai state funds invest 1 billion yuan in Zhipu
At its Shanghai ecosystem conference, Zhipu announced 1 billion yuan in new funding from Pudong Venture Capital and Zhangjiang Group, following state-backed investments from Beijing, Chengdu, Hangzhou and Zhuhai earlier in 2025. It also open-sourced GLM-4.1V-Thinking and launched a MaaS agent platform.
Zhipu rebrands to Z.ai and releases GLM-4.5 open-source model
Zhipu AI rebranded internationally as Z.ai while releasing GLM-4.5 and GLM-4.5-Air under the MIT license. GLM-4.5, trained on 15 trillion tokens with 355 billion parameters and 32 billion active, featured a hybrid reasoning mode and up to 96K output tokens. The launch report also described expansion with offices in the US, UK and France. A separate November report later documented a tenfold increase in overseas paid users.
GLM-4.5 immediately refuses a query about Jiangyou police
During protests in Jiangyou, Sichuan over the police response to a bullying case, China Media Project tested the Chinese phrase 'Jiangyou police.' Zhipu's GLM-4.5 immediately said it could not answer related questions. The report also described refusals or avoidance when asking models about the bullying incident. These tests illustrated restrictions on access to sensitive breaking-news information; they did not establish that every possible query would be refused.
Mandatory AI content labeling takes effect in China
China's Measures for Labeling AI-Generated Synthetic Content took effect, requiring all generative AI content to carry both explicit labels (visible 'AI-generated' notices) and implicit labels (digital watermarks). Zhipu implemented compliance measures including explicit labels on Qingyan responses and hidden digital identifiers on all text, images, and videos.
RSF documents systematic censorship in Chinese AI chatbots
Reporters Without Borders published an investigation of state propaganda and censorship in Chinese AI chatbots, testing DeepSeek, Ernie and Qwen across around 30 topics and over 100 prompts. It found refusals and responses aligned with official narratives across several languages. These were findings about other Chinese chatbot providers, offering sector context for Zhipu's documented censorship obligations.
Zhipu says restructuring affected more than ten people; former staff claim nearly 100
Responding to reports of layoffs and department closures, Zhipu told First Financial that its product and R&D departments were operating normally and that an organizational adjustment involved more than ten people. Former employees had told reporters that nearly 100 were affected, some signing out the same day.
GLM-4.7 open-sourced as coding-focused agent model
Z.ai released GLM-4.7 on December 22 as an open model for coding and multi-step tool workflows. Its Preserved Thinking capability retains reasoning across turns. Z.ai reported 87.4 on the interactive tool-use benchmark τ²-Bench, the highest reported open-model result at the time. Contemporary Techloy coverage described $3/month cloud access, compared with premium US coding subscriptions at $200/month, while public weights allowed local deployment. Subsequent US developer uptake is covered by the separately dated TechBuzz evidence item.
China proposes draft rules for human-like interactive AI services
On December 27, the CAC released draft rules for public-facing human-like interactive AI services. The proposal would prohibit harmful emotional manipulation, require assessment of user dependence and intervention when addiction is detected, mandate guardian consent for minors' emotional-companionship use, and require security assessments above 1 million registered users or 100,000 monthly active users. These proposed obligations would be relevant to qualifying AI-companion services, including any offered by Zhipu's Lingxin affiliate, rather than imposing an immediate legal duty on Qingyan.
Z.ai IPO on Hong Kong Stock Exchange raises $558 million
Z.ai, listed as Knowledge Atlas Technology, became the first major Chinese LLM developer to go public on HKEX. The IPO raised about HK$4.3 billion ($558 million) at HK$116.20 per share, and the shares closed at HK$131.5. The company earmarked 70% of proceeds for R&D. CGTN also reported an opening price of HK$120, market value above HK$57 billion and 1,159-times public-offering oversubscription, while attributing to Zhipu the claim to be the world's first public company focused on AGI foundation models.
Coding plan sign-ups restricted to 20% amid computing constraints
On January 21, Z.ai said it would limit new GLM Coding Plan sign-ups to 20% of existing levels, an 80% reduction, and crack down on malicious use. SCMP reported performance issues for existing users amid computing constraints. Later reporting described January refund demands and response delays and rate limits in February. The global appeal for computing partners occurred in February, rather than as part of the January sign-up announcement.
GLM-5 release accompanied by Coding Plan and API price increases
Following GLM-5's February 11 release, Z.ai raised Coding Plan prices by at least 30% on February 12 while leaving existing subscribers on their original rates and removing first-purchase discounts. TrendForce reported overseas Coding Plan increases of 30–60% and overseas API increases of 67–100%; these are different products, not domestic versus international subscription bands. GLM-5 has 744 billion total and 40 billion active parameters. Zhipu shares rose roughly 30% on February 12. The report called this the first significant domestic LLM price hike of 2026, while noting that some providers had already raised API prices in 2025.
PNAS study documents systematic political censorship in Chinese LLMs
A peer-reviewed PNAS Nexus study by Stanford and Princeton researchers tested nine models, including ChatGLM and three other China-originating models, with 145 political questions. With Chinese prompts, ChatGLM refused about 10%, the lowest China-originating refusal rate, versus about 36% for DeepSeek and 0–2.8% for non-China models. China-originating models generally gave shorter and less accurate answers. The authors explicitly cautioned that the observational study did not establish a causal link between regulation and model behavior or identify alignment training as the mechanism.
Zhipu apologizes for GLM-5 Coding Plan rollout and offers refunds
Zhipu admitted three mistakes after the GLM-5 launch: insufficiently transparent rules, a slow staged rollout and a poorly designed upgrade mechanism for existing users. With demand exceeding capacity, Max users got full access, Pro users faced peak-hour limits and Lite users were to get off-peak access after Chinese New Year; affected Lite and Pro users could cancel and request refunds before March 6.
Shares tumble 23% on computing woes and user complaints
Z.ai's shares fell nearly 23% on Monday, February 23, amid computing-capacity concerns and continuing complaints about service quality. The company had issued a global appeal for computing partners the previous week. Users in China and overseas reported Coding Plan response delays and rate limits, as documented by the separately cited Silicon report. Investing.com subsequently reported an intraday rebound of more than 20% on February 24. These reports describe the circumstances around the stock moves without establishing that complaints alone caused them.
AutoClaw launched as domestic OpenClaw alternative
Z.ai launched AutoClaw on March 10, billing it as China's first one-click-install local version of OpenClaw with more than 50 pre-installed skills and its Pony-Alpha-2 model. Shares closed almost 13% higher, extending gains since the January IPO to more than 300%. Yicai reported that AutoClaw's installation process included measures to avoid exposed public-network ports and to restrict important operations. Bloomberg reported on March 11 that certain employees at state-run banks and some government agencies were barred from installing OpenClaw on office computers; other notices required prior approval.
First post-IPO results: 2025 revenue up 131.9%, losses up 59.5%
Zhipu reported 2025 revenue of 724.33 million yuan, up 131.9% year on year and below analyst estimates. Total losses rose 59.5% to 4.72 billion yuan, driven by R&D spending that rose 44.9% to 3.18 billion yuan.
GLM-5.1 released with another round of API price increases
Zhipu released the open-source GLM-5.1 model. OpenRouter showed GLM prices up about 10%, with cloud call prices 8% to 17% above the earlier GLM-5 Turbo, at least the second Zhipu price increase of 2026 after February's Coding Plan rise and covering more billing dimensions including cloud input tokens.
China finalizes human-like interactive AI rules, effective July 15
The CAC and four other bodies promulgated interim measures for anthropomorphic interactive AI services, finalizing the December 2025 draft, with effect from July 15, 2026. The rules cover continuous emotional-interaction services such as companionship and emotional support; non-emotional customer service, knowledge Q&A, work assistance, education and research services are excluded.
2025 annual report shows 157 million yuan pay for executive director
Zhipu's 2025 annual report listed 1,094 employees and 1.363 billion yuan in total compensation costs including share-based payments. Executive director Liu Debing's total remuneration was 157 million yuan and Zhang Peng's 68.43 million yuan, both including large share-based payments.
GLM-5.2 announced with one-million-token context and MIT license
Zhipu announced GLM-5.2 as its latest flagship model, supporting context lengths of up to one million tokens, and said it would be released as open source under the MIT license.
CAC campaign report cites Zhipu's multimodal moderation upgrade
In its first-phase report on the national campaign against AI application problems, the CAC said Zhipu had built a multimodal review model and added cross-review for complex multi-turn dialogue cases such as intent reversals, citing it among platforms meeting their management responsibilities.
Zhipu completes HK$31.41 billion follow-on share placement
Zhipu completed a placement of new H shares on the Hong Kong Stock Exchange on July 13, 2026, with gross proceeds of about HK$31.41 billion (about US$4 billion). The filing said the placement would broaden its shareholder base and support expansion of its general-purpose AI business.
Zhipu reportedly builds 1-GW computing centre and acquires XCore Sigma
SCMP reported, citing people familiar with the matter, that Zhipu had built a 1-gigawatt AI computing centre to train and deploy GLM models and had completed the acquisition of XCore Sigma, an infrastructure software developer spun off from the Chinese Academy of Sciences.
Qingyan PC client launched with local-folder analysis
Zhipu launched a Qingyan PC client combining Chat and AgentMore modes with context carried between them. It can mount local folders so the assistant tracks file changes and analyzes across files.
Points-based Coding Plan raises new-user monthly fees by up to 261%
Zhipu launched a new GLM Coding Plan and opened full subscriptions after months of limited daily sales. New Lite, Pro and Max monthly fees of 118, 538 and 1,078 yuan replaced 49, 149 and 469 yuan, increases of about 141%, 261% and 130%. Usage limits moved from prompt counts to points covering input, output, cache-hit tokens and MCP calls, and it was unclear whether allowances rose; existing V2 subscribers kept their prices.
H1 2026 revenue up nearly 400%, led by API services
Z.ai reported first-half 2026 revenue of about 954 million yuan, up nearly 400% year on year. Open-platform and API revenue was about 825 million yuan, up 2,735.7%.
Zhipu completes HK$714 share placement and RMB 20.14 billion convertible bond issue
Zhipu allotted 21,965,000 new H shares at HK$714 each to at least six placees, about 4.50% of its enlarged share capital. It also completed an issue of convertible bonds with an aggregate principal of RMB 20.14 billion.
Employee share awards granted over 1.22 million H shares
Z.ai granted awards underlying 1,222,934 H shares to employee participants under its 2026 Share Award Scheme, subject to acceptance.
Evidence (64 citations)
D1: User Value Erosion
D2: Business Customer Exploitation
D3: Shareholder Extraction
D4: Lock-in & Switching Costs
D5: Twiddling & Algorithmic Opacity
D6: Dark Patterns
D7: Advertising & Monetization Pressure
D8: Competitive Conduct
D9: Labor & Governance
D10: Regulatory & Legal Posture
Scoring Log (6 entries)
38→39. Since Feb 2026: GLM-5 apology and refunds (Feb 22), 2025 results with losses of 4.72B yuan and no dividend, GLM-5.1 with ~10% API price rise and CEO-reported 83% Q1 API price increase, final human-like interactive AI rules (Apr 10, effective Jul 15; exclude Q&A/work assistants), GLM-5.2 MIT (Jun), CAC campaign citing Zhipu moderation (Jul 6), HK$31.41B placement (Jul 13), Qingyan PC client (Jul 25), points-based Coding Plan raising new-user fees ~130-261% (Jul 31), H1 2026 revenue up ~400%, HK$714 placement plus RMB20.14B convertible bonds (Sep 18). D2 3→5 (event: 2026 API/Coding Plan increases in Apr and Jul on top of Feb, CEO-reported 83% Q1 API price rise; grandfathering and refunds keep it below 6). D7 3→4 (event: Jul 31 points-based Coding Plan repricing up to 261% for new users, harder tier comparison; still no documented ads). D8 4→3 (recalibration: small acquisitions, open MIT releases, no antitrust matters; criteria 2-3 fit). D10 6→5 (recalibration: Entity List plus a mixed domestic compliance record; no lobbying, litigation against critics or decree violations required for 6-7). Eras: all five re-dated to inflection events (2019-06-01→2019-06-11 founding; 2022-08-01→2022-08-04 GLM-130B; 2023-08-01→2023-08-31 Qingyan launch, relabeled 'ChatGLM Consumer Launch'→'Qingyan Public Launch'; 2025-01-01→2025-01-16 Entity List; current era 2026-02-20 (assessment date)→2026-01-08 IPO). All eras re-scored: era 2 D10 4→2 and era 3 lowered (22→18) as the timeline shows an improving product and developer-friendly pricing then. Removed two synthetic-analysis evidence items (Finterra, PredictStreet). Added 23 timeline events (9 historical) and 16 evidence items. Category AI fine.
Checked 5 removed/trimmed candidates: 0 restored, 3 partly restored, 2 confirmed removed, 0 already present. PARTLY RESTORED timeline 2: Carnegie (2023-07-10) supports among the earliest and most detailed for the broader Chinese AI framework, not an absolute global ranking of this provision; timeline 6: Economic Observer via The Paper (2023-12-11) reports over 10 million ChatGLM-6B-series downloads, not the first model alone; timeline 39: Bloomberg via The Edge (March 11 report in March 12 issue) confirms some bank/agency installation bans and approval notices, not AutoClaw launch causality. CONFIRMED REMOVED timeline 14: Center for Data Innovation (2024-12-12), Turing Post (2024-05-04), TechCrunch (2023-10-20) and Beike Finance via Xinhua Finance (2024-12-17) support the corrected workforce, cumulative fundraising and later user count already retained, not the asserted October 2023 $350M single-round-to-workforce causality or 25M users then. CONFIRMED REMOVED timeline 16: original Caproasia page and USCC 2024 Chapter 5 establish participation/minority investment, not a Series C label, a lead role or direct sovereign-project causality. All original URLs attempted; four live pages read, Bloomberg launch blocked (403) and recovered through China Daily's matching agency report. Explicit Wayback attempts for all five originals returned no content and were not used as negative evidence. Used 9 WebSearch calls total (per candidate: 2,2,1,1,3), with exact phrases, named sources and date drift considered; session source extracts and candidate decisions retained in work/restore-decisions.json. No removals or score changes.
Checked 82 items: original68timeline/evidence/milestones +3alternatives +11new supporting evidence items preserving original specifics. 29 verified, 34 corrected (15 date-only), 19 re-sourced, 0 removed. No score changes. Nine strict invented details: June 2024 GLM-4 model weights released under MIT — Weights use a custom Model License; code uses Apache 2.0. Source: https://github.com/zai-org/GLM-4/blob/main/README_20240605.md; Zhipu plus ten subsidiaries in the January 2025 rule — Ten related entities in total: parent plus nine. Source: https://www.federalregister.gov/documents/2025/01/16/2025-00704/addition-of-entities-to-and-revision-of-entry-on-the-entity-list; January 2025 special Footnote 4 designation — Final rule assigns ordinary EAR license controls; no Footnote 4. Source: https://www.federalregister.gov/documents/2025/01/16/2025-00704/addition-of-entities-to-and-revision-of-entry-on-the-entity-list; 67-100% increase applied to international Coding Plan subscriptions — 67-100% is overseas API fees; Coding Plan band is 30-60%. Source: https://www.trendforce.com/news/2026/02/16/news-rising-costs-and-demand-drive-chinas-llm-price-jump-zhipu-glm%E2%80%915-hikes-30-in-first-2026-increase/; GLM-5 has 44B active parameters — Official card gives 40B active of 744B total. Source: https://huggingface.co/zai-org/GLM-5; Consumer tariff of 9.9 yuan/month or 0.1 yuan/use — JUNK-ONLY: exact schedule traces only to SEO directory; official iOS listing describes membership, SVIP and credit purchases (https://apps.apple.com/cn/app/id6450893458). Source: https://www.nettsz.com/chatglm; Consensus breakeven estimate of 2027-2028 — JUNK-ONLY: exact band traces to Finterra synthetic analysis; authentic earlier tentative CEO estimate was 2026/27 (https://www.turingpost.com/p/zhipu). Source: https://markets.financialcontent.com/wral/article/finterra-2026-2-10-deep-dive-knowledge-atlas-hkex-2513-the-glm-architect-and-chinas-agi-race; R&D consistently exceeded revenue by 8:1 through 2024 — 2023 R&D/revenue about4.2x; 2024 about7x; H1 2025 about8.35x. Source: https://www1.hkexnews.hk/listedco/listconews/sehk/2025/1230/2025123000017.pdf; Tang Jie said money should not be a hindrance on the road to AGI — Profile attributes the internal saying about fundraising/earnings to CEO Zhang Peng, not Tang Jie. Source: https://www.turingpost.com/p/zhipu. Unsupported claims: 2022 algorithm-regulation global superlative; Prosperity7 round Series C label; Funding directly enabled sovereign infrastructure projects; AutoClaw launched in response to OpenClaw bans; RMB15B user-acquisition spending; Absolute conversation-history nonexportability; Collegial governance culture inferred from university origins; 84.5% as a government/SOE customer share. Prose claim ledger and full search logs retained in work/claim-decisions.json, work/prose-findings.json, work/prose-open-findings.json and work/timeline-first-findings.json.