China's Hundred Model War and the Race That Reshaped Global AI

The term “Hundred Model War” (百模大战) describes the explosive proliferation of large language models in China that began in early 2023, triggered directly by OpenAI’s release of ChatGPT in November 2022. Within months, more than one hundred domestic large models had been announced across the country. By July 2025, Chinese institutions accounted for roughly 1,509 of the world’s approximately 3,755 publicly disclosed LLMs. The phrase itself, which deliberately echoes the Chinese Communist Party’s celebrated Hundred Regiments Offensive (百团大战) of 1940, entered public discourse around March 2023 and was formally recognized when the National Language Resources Monitoring and Research Center included it among its Top Ten New Chinese Media Buzzwords of 2023. What followed was not merely a technology race but a coordinated national mobilization involving state policy, private capital, and sweeping regulatory architecture—one that, by early 2026, had produced genuine frontier capabilities while also confirming predictions of brutal consolidation.

The regulatory scaffolding for this mobilization rests on the State Council’s New Generation Artificial Intelligence Development Plan (新一代人工智能发展规划), issued in July 2017, which committed China to achieving world-leading status in AI theory, technology, and applications by 2030. That framework was operationalized for generative AI through the Interim Measures for the Management of Generative AI Services (生成式人工智能服务管理暂行办法), jointly issued by seven agencies led by the Cyberspace Administration of China on July 10, 2023 and effective from August 15, 2023. These Interim Measures, spanning five chapters and twenty-four articles, made China the first country with binding regulations specifically targeting generative AI. They require providers of public-facing generative AI services to complete a security assessment and algorithm filing (备案) with the CAC, mandate adherence to core socialist values, and impose obligations regarding training data legality, personal data consent, and AI-generated content labeling. Research and internal use were explicitly excluded from the compliance requirements, allowing companies to continue developing models freely while regulating only their public deployment. By the end of 2025, 748 generative AI services had completed national filing and an additional 435 had completed local-level registration, for a total of 1,183 approved services. Subsequent regulatory developments included the AI Content Labeling Measures effective September 2025, the State Council’s landmark AI Plus Action Plan (国发〔２０２５〕11号) issued in August 2025 establishing phased integration targets through 2035, and amendments to the Cybersecurity Law passed on October 28, 2025 and effective January 1, 2026, which for the first time brought artificial intelligence into national statute. A comprehensive standalone AI law was placed on China’s 2025 legislative agenda but removed in May of that year, signaling a deliberate preference for incremental regulation over a single omnibus framework.

Baidu was the first major Chinese technology company to unveil a ChatGPT competitor. Its product, officially named ERNIE Bot in English and 文心一言 (Wenxin Yiyan) in Chinese, debuted in an invited demonstration on March 16, 2023. That initial showing was reportedly pre-recorded rather than live, and it caused Baidu’s stock to drop ten percent intraday before recovering the following day after more favorable analyst reviews. The public launch came on August 31, 2023, when ERNIE Bot was among the first eleven generative AI products to receive CAC regulatory approval under the newly effective Interim Measures. The underlying model family traces a long lineage. Baidu’s original ERNIE (Enhanced Representation through Knowledge Integration) architecture dates to 2019, with the ERNIE 3.0 Titan variant reaching 260 billion parameters in 2021. Subsequent iterations arrived in rapid succession: ERNIE 3.5 in June 2023 with plugin integration; ERNIE 4.0 in October 2023, positioned as a direct GPT-4 competitor; ERNIE 4.0 Turbo in June 2024, optimized for inference speed; ERNIE 4.5 alongside a dedicated deep-reasoning model called ERNIE X1 on March 16, 2025; and ERNIE 5.0, unveiled at Baidu World 2025 on November 13, 2025, as a natively omni-modal model jointly processing text, images, audio, and video through a 2.4-trillion-parameter mixture-of-experts architecture, with a stable update released on January 10, 2026.

Baidu reported cumulative registered users reaching 100 million by December 2023, 200 million by April 2024, and 300 million by June 2024, though these figures reflect registrations rather than monthly engagement. By December 2025, the ERNIE Assistant reported approximately 200 million monthly active users. ERNIE Bot was integrated across Baidu’s ecosystem, powering AI-enhanced search results, driving the Qianfan model-as-a-service cloud platform (which generated over 30 billion yuan in AI cloud revenue in 2025), embedded in Samsung Galaxy devices for the China market beginning January 2024, and underpinning productivity tools, digital humans for livestream commerce, and the Apollo Go autonomous ride-hailing service operating across twenty-six cities. In September 2024, the Chinese-language chatbot was rebranded from 文心一言 to 文小言 (Wenxiaoyan) as part of a repositioning toward search assistance. On April 1, 2025, under competitive pressure from DeepSeek and other free alternatives, Baidu made ERNIE Bot entirely free for all users, abandoning its prior subscription model for ERNIE 4.0 access.

Alibaba’s earlier M6 model—the Multi-Modality to Multi-Modality Multitask Mega-transformer, developed by DAMO Academy (达摩院) and published in March 2021—was a multimodal pre-training model designed for cross-modal tasks such as text-guided image generation. It scaled impressively from 10 billion to an announced 10 trillion parameters by November 2021, but M6 was not a conversational large language model. It was a research milestone focused on Chinese multimodal understanding, applied primarily in product description generation and design within Alibaba’s e-commerce ecosystem. It is a predecessor to, but architecturally and functionally distinct from, Alibaba’s subsequent chatbot and LLM family. That family is Tongyi Qianwen (通义千问), launched in beta with invitation-only access in April 2023 and opened to the public in September 2023 after receiving CAC regulatory clearance. The product was renamed simply Tongyi (通义) in May 2024, while the model brand became Qwen internationally. DAMO Academy, established in October 2017 as Alibaba Group’s global research institute, conducted the foundational research, while Alibaba Cloud served as the commercialization and distribution arm through its DashScope API platform and the ModelScope community for open model hosting.

The Qwen model family evolved rapidly. Qwen-7B was released as open weights on Hugging Face in August 2023. The Qwen2 series launched in June 2024, and Qwen2.5 followed in September 2024 spanning models from 0.5 billion to 72 billion parameters. The Qwen3 family arrived on April 28, 2025, comprising dense models from 0.6 billion to 32 billion parameters and mixture-of-experts models up to 235 billion total parameters with 22 billion active, all trained on 36 trillion tokens across 119 languages and released under the Apache 2.0 license. By early 2026, the series had progressed through Qwen3-Max, Qwen3-Omni, and Qwen3.5 alongside Qwen3.5-Plus. Alibaba’s open-weight strategy proved transformative: by January 2026, Qwen models had been downloaded over 700 million times on Hugging Face, surpassing Meta’s Llama in October 2025 to become the world’s most downloaded model family, with more than 180,000 community-built derivative models. Singapore’s National AI Programme adopted the Qwen architecture for its Southeast Asian language model in November 2025. Alibaba committed $53 billion over three years to cloud and AI infrastructure, and by the first half of 2025 its AI products had grown at triple-digit rates for eight consecutive quarters, with the company claiming approximately one-third of China’s AI cloud market.

Tencent’s large language model is called Hunyuan (混元). The company officially unveiled it at its Global Digital Ecosystem Conference in September 2023 as a model with hundreds of billions of parameters. The consumer-facing AI assistant built on Hunyuan, called Yuanbao (元宝), launched on app stores on May 30, 2024. The model family progressed through increasingly capable versions: Hunyuan Turbo with a heterogeneous mixture-of-experts architecture in September 2024; the open-sourced Hunyuan-Large, with 389 billion total parameters and 52 billion active, in November 2024, at that time the largest open-source Transformer-based MoE model; Hunyuan T1 for deep reasoning in March 2025; and Hunyuan 2.0 in December 2025 with 406 billion total parameters, 32 billion active, and a 256,000-token context window. During its fourth-quarter 2025 earnings call in March 2026, Tencent announced that Hunyuan 3.0 was in internal testing with a planned April 2026 launch. Tencent’s integration strategy leveraged its unmatched distribution through WeChat. In April 2025, Yuanbao was embedded directly as a contact within WeChat, and by mid-2025 the AI assistant was handling over ninety percent of question-based queries within WeChat’s search function. Integration extended across QQ Browser, Tencent Meeting, Tencent Docs, and the ima AI workbench. By the second quarter of 2025, Yuanbao had reached approximately 41.6 million monthly active users and ranked as the third most popular AI application in China by daily active users. Tencent simultaneously pursued a dual strategy of developing its own models while investing in external AI startups including Moonshot AI, Baichuan Intelligence (百川智能), MiniMax, and Zhipu AI, and notably integrated DeepSeek models alongside Hunyuan across multiple products beginning in February 2025.

Huawei’s Pangu model, named after the mythological creator of the world in Chinese cosmology, launched with PanGu-α in July 2021 as a 200-billion-parameter autoregressive Chinese language model, at the time described as the world’s largest Chinese NLP model. Developed by Huawei’s Noah’s Ark Lab, it was trained on Huawei’s own Ascend 910 AI processors using the MindSpore (昱思) framework, an open-source deep learning platform analogous to Google’s TensorFlow and Meta’s PyTorch. MindSpore was announced in August 2019 alongside the Ascend 910 chip and open-sourced in March 2020, supporting cloud, edge, and device deployment, optimized for Huawei’s Ascend processors through the CANN (Compute Architecture for Neural Networks) software stack. The Pangu family expanded through PanGu-Σ in 2023, a 1.085-trillion-parameter sparse mixture-of-experts model; Pangu 3.0 in July 2023, introducing a three-tier architecture for enterprise customization; Pangu 5.0 in June 2024 with variants ranging from one billion to over one trillion parameters integrated with HarmonyOS; and Pangu 5.5 in June 2025, a 718-billion-parameter deep-reasoning MoE model with 256 experts deployed across more than 500 scenarios in over 30 industries. In June 2025, Huawei also open-sourced the openPangu family, though controversy arose in July 2025 when observers on GitHub noted structural similarities between the openPangu Pro MoE and Alibaba’s Qwen architecture.

What fundamentally distinguishes Huawei from other participants in the Hundred Model War is its vertical integration across the entire AI technology stack: from chips through frameworks to models to cloud platforms to applications. Under the pressure of U.S. sanctions that cut off access to TSMC’s advanced fabrication, Huawei shifted Ascend production to SMIC’s 7nm DUV multi-patterning process. The current flagship Ascend 910C, a dual-chiplet design with 96 gigabytes of HBM2e memory, delivers roughly 780 to 800 teraflops in FP16/BF16—approximately one-third the throughput of Nvidia’s B200 but critically available to Chinese AI developers when Western alternatives are not. Huawei’s strategic focus has remained enterprise and industrial AI rather than consumer chatbots, with Pangu models deployed in government, finance, manufacturing, mining, meteorology, and energy applications. In this sense, Huawei functions less as an LLM competitor and more as the domestic infrastructure provider enabling other Chinese companies to train and deploy their models under sanctions constraints.

ByteDance’s AI chatbot is called Doubao (豆包), literally “bean-paste bun.” The underlying model was initially named Skylark (云雀) when ByteDance launched it through its Volcano Engine (火山引擎) cloud platform on August 6, 2023, and has since evolved through the Seed research team’s model series into Doubao-1.5-pro in January 2025 and Doubao-Seed-2.0 on February 14, 2026. Doubao’s trajectory illustrates how distribution advantages can prove decisive in consumer AI. ByteDance leveraged its existing ecosystem—Douyin, Toutiao, Feishu, and over fifty other internal products—to drive adoption at extraordinary scale. On May 15, 2024, Volcano Engine launched nine enterprise Doubao models at prices 99.3 percent below the industry average, triggering a devastating price war across China’s LLM market. By the end of March 2025, Doubao was processing 12.7 trillion tokens daily, a 106-fold year-over-year increase. Daily active users surpassed 100 million by December 2024, and by February 2026 QuestMobile data showed 227 million users in the native AI generation category, making Doubao China’s most popular AI application by a margin of roughly 100 million users over its nearest competitor. The Seed 2.0 model ranked sixth on the LMSYS Chatbot Arena for text and third on the Vision Arena as of its release.

SenseTime, founded in October 2014 by Tang Xiao’ou, a professor at the Chinese University of Hong Kong, listed on the Hong Kong Stock Exchange in December 2021 after overcoming a last-minute disruption when the U.S. Treasury placed it on the Non-SDN Chinese Military-Industrial Complex Companies list on December 10, 2021—deliberately timed to Human Rights Day, the same day SenseTime had planned to price its IPO. The company had already been added to the U.S. Commerce Department’s Entity List in October 2019 as part of sanctions on twenty-eight Chinese entities for alleged involvement in human rights abuses against Uyghurs in Xinjiang. Tang Xiao’ou died on December 15, 2023, at age fifty-five, and leadership passed to co-founder and CEO Xu Li. SenseTime’s government ties are deep and structural. It was designated one of China’s national AI champions in 2018, named a National Open Innovation Platform for next-generation AI in intelligent vision, and holds a golden share controlled by the China Internet Investment Fund, a state-owned enterprise under the CAC, making SenseTime partly state-owned by design. Its generative AI platform, SenseNova (日日新, Rì Rì Xīn), has iterated aggressively from its initial 2023 release through version 6.5, unveiled at the World AI Conference in July 2025. The company has struggled financially, posting persistent net losses, though it narrowed its adjusted net loss by fifty percent in the first half of 2025 to 1.16 billion yuan as generative AI revenue grew 72.7 percent year-over-year to constitute seventy-seven percent of total revenue. SenseTime is building domestic chip compatibility through partnerships with Huawei, Hygon, Cambricon, Biren, and Moore Threads, operating what it calls a compute mall for heterogeneous AI infrastructure.

Zhipu AI (智谱AI) was founded on June 11, 2019, by Tang Jie and Li Juanzi, both professors at Tsinghua University’s Department of Computer Science, as a spinoff of the university’s Knowledge Engineering Group. Tang Jie, an IEEE, ACM, and AAAI Fellow who also served as a technical lead in the Beijing Academy of Artificial Intelligence’s WuDao project, is the company’s chief scientist and controlling shareholder. Zhang Peng serves as CEO. The company’s General Language Model (GLM) architecture has progressed from GLM-10B in 2021 through the landmark GLM-130B in 2022, a 130-billion-parameter bilingual model that was open-sourced, and the widely adopted ChatGLM-6B series beginning in 2023, which accumulated over 30 million global downloads and could run on consumer-grade GPUs. Subsequent versions included GLM-4 in 2024, GLM-4.5 in July 2025 as a 355-billion-parameter open-source MoE model, and GLM-5 in 2026 as a unified multimodal model at 744-billion-parameter scale. In July 2025 the company rebranded internationally as Z.ai. Zhipu AI’s funding journey reflects the arc of the Hundred Model War itself. The company struggled to raise capital in its early years, with Chinese venture capitalists skeptical of large model potential until ChatGPT’s November 2022 debut transformed sentiment overnight. A 2023 Series B round of 2.5 billion yuan, led by Meituan with participation from Alibaba, Tencent, and Xiaomi, established Zhipu as a leading AI Tiger. A subsequent $400 million round in May 2024 valued the company at approximately $3 billion. Zhipu completed its initial public offering on the Hong Kong Stock Exchange on January 8, 2026, raising $558 million—the largest IPO for an AI foundation model company to date—and saw its stock surge 173 percent in the first month of trading. On January 16, 2025, the U.S. Commerce Department added Zhipu AI to its Entity List, cutting off access to Nvidia chips and accelerating the company’s pivot to training on domestic hardware including Huawei Ascend, Cambricon, and Moore Threads processors.

DeepSeek (深度求索) was founded in mid-2023 by Liang Wenfeng, a Zhejiang University graduate who co-founded the quantitative hedge fund High-Flyer (幻方量化). DeepSeek is unique among China’s AI companies in that it has accepted no external venture funding, operating entirely on High-Flyer’s capital. Liang, who holds approximately eighty-four percent ownership through shell corporations, built a research-focused organization in Hangzhou that prioritized fundamental capability over commercial applications. DeepSeek’s impact crystallized on January 20, 2025, with the release of DeepSeek-R1, a 671-billion-parameter reasoning model with 32 billion active via mixture-of-experts, released under the MIT License. R1 achieved performance parity with OpenAI’s o1 across mathematical, coding, and reasoning benchmarks. The innovations were substantive: reinforcement learning without supervised fine-tuning, demonstrated through the R1-Zero variant; the Group Relative Policy Optimization algorithm; and multi-head latent attention for memory efficiency. One week later, on January 27, 2025—dubbed DeepSeek Monday—the DeepSeek app surpassed ChatGPT as the top free application on the U.S. iOS App Store. Nvidia lost $600 billion in market capitalization in a single trading day, the largest single-company decline in American stock market history, as investors questioned whether the massive GPU buildout was overbuilt. Over one trillion dollars was erased from U.S. technology stocks. The event was widely characterized as a Sputnik moment for artificial intelligence, and Liang Wenfeng was subsequently invited to symposia hosted by both Premier Li Qiang and President Xi Jinping.

The broader ecosystem around these principal players produced its own significant developments. Moonshot AI (月之暗面), founded in March 2023 by Yang Zhilin, a Tsinghua and Carnegie Mellon alumnus, built the Kimi chatbot with breakthrough long-context processing capabilities, initially handling 200,000 Chinese characters and later two million. By early 2026 the company had released Kimi K2.5, a one-trillion-parameter open-weight MoE model with agentic capabilities. MiniMax, founded in December 2021 by Yan Junjie, developed the Talkie companion chatbot, ranking among the top entertainment apps in the United States, while simultaneously pushing its M2.5 model to the number-one position globally by token consumption on the OpenRouter platform in February 2026. iFlytek (科大讯飞), a partially state-owned company founded in 1999 and listed on the Shenzhen Stock Exchange, released its Spark (讯飞星火) model series beginning in May 2023 and achieved a notable milestone in April 2025 when its Xinghuo X1, a 70-billion-parameter deep-reasoning model co-developed with Huawei, became the first major Chinese LLM trained entirely on domestic Ascend hardware.

Perhaps the most telling development was the trajectory of 01.AI (零一万物), founded in March 2023 by the prominent AI figure Kai-Fu Lee. After an initial burst of activity—its Yi-34B topped Hugging Face leaderboards in November 2023 and the company reached unicorn status within eight months—01.AI announced in March 2025 that it would stop pre-training large language models entirely, forming a joint lab with Alibaba Cloud and pivoting to enterprise solutions built on top of DeepSeek. This effective exit from the frontier model race by one of the most well-resourced and well-connected startups confirmed the consolidation that industry observers had long predicted. Baichuan Intelligence (百川智能), founded by former Sogou CEO Wang Xiaochuan, similarly pivoted toward specialized healthcare and finance verticals rather than continuing to compete at the general-purpose frontier.

The competitive dynamics of the Hundred Model War cannot be understood apart from the escalating U.S. semiconductor export controls that constrained Chinese access to advanced AI hardware. The initial controls of October 7, 2022, barred export of Nvidia’s A100 and H100 chips to China. When Nvidia responded with compliant variants, the A800 and H800, a major update on October 17, 2023 closed that loophole by shifting to performance-density thresholds, controlling the replacement chips as well. Nvidia then designed the H20, specifically engineered to fall below the new thresholds, but this too was restricted in April 2025 when the Trump administration imposed an indefinite license requirement, prompted partly by evidence that H20 chips had been used to train DeepSeek’s models. In May 2025, the Bureau of Industry and Security issued guidance declaring Huawei’s Ascend 910B, 910C, and 910D chips presumptively produced in violation of U.S. export controls, establishing a global prohibition on their use without BIS authorization. In a reversal reflecting the complexity of the policy landscape, the Trump administration in December 2025 approved sales of Nvidia’s H200 to China—the most powerful chip ever authorized for Chinese export—driven by concerns that continued restrictions were accelerating Huawei’s competitive position as a domestic alternative.

Despite these constraints, approximately seventy-five percent of chips powering AI model training in Chinese data centers still ran on Nvidia’s CUDA platform as of 2025, reflecting both existing inventory and the flow of chips through third countries. China responded with domestic alternatives—Huawei’s Ascend series, Baidu’s Kunlun chips, Alibaba’s T-Head processors, and emerging competitors like Cambricon and Moore Threads—and with retaliatory export controls on critical minerals including gallium, germanium, antimony, and graphite. The chip restrictions unquestionably increased costs and complicated training at scale, but they also galvanized domestic semiconductor development and, paradoxically, may have contributed to the efficiency innovations that made DeepSeek R1 possible.

By early 2026, China’s LLM landscape had undergone the consolidation that Baidu CEO Robin Li publicly warned about at the World AI Conference in July 2024, when he declared that the Hundred Model War had caused a huge waste of resources, especially computing power. The market divided into distinct tiers. A handful of organizations continued competing at the global frontier: DeepSeek as a research-heavy infrastructure provider, Alibaba’s Qwen as the dominant open-weight ecosystem, ByteDance’s Doubao as the consumer distribution leader with 227 million users, and Moonshot AI and MiniMax as agile frontier labs with both model capability and consumer traction. The established technology giants—Baidu, Tencent, and Huawei—retained significant positions through ecosystem integration and cloud infrastructure but found themselves increasingly incorporating external models like DeepSeek alongside their proprietary offerings. A second tier of companies pivoted away from frontier competition entirely, with 01.AI abandoning pre-training, Baichuan retreating to vertical specialization, and Zhipu AI focusing on commercial applications and its landmark IPO.

The price war that DeepSeek triggered in May 2024, when its V2 model launched at $0.14 per million input tokens, five to ten times below prevailing rates, cascaded through the entire market. Alibaba slashed Qwen pricing from $1.10 to $0.07 per million tokens. ByteDance launched Doubao lite at $0.04. By January 2025, average Chinese LLM API prices had fallen ninety-two percent from their May 2024 levels, and the collapse rippled globally, forcing OpenAI, Anthropic, and Google into their own pricing reductions. The Jevons paradox proved operative: as the cost of AI reasoning plummeted, total demand for AI services exploded rather than contracted, and Nvidia recovered to reach a $5 trillion market capitalization by October 2025.

The Hundred Model War produced outcomes that few predicted in early 2023. China did not merely replicate Western models at a lag; it developed distinctive competitive advantages in efficiency, open-weight distribution, and price-driven market expansion. The regulatory framework, while constraining in some respects, provided a structured pathway for deployment that prevented the chaotic unregulated proliferation seen in some other markets. By early 2026, with over 515 million generative AI users in China representing 36.5 percent population penetration, and with Chinese open-source models accounting for roughly thirty percent of global AI downloads, the Hundred Model War had evolved from a domestic competition into a force reshaping the global artificial intelligence landscape. The question was no longer whether China could build competitive large language models, but whether the consolidation phase would produce durable companies capable of sustaining the enormous capital requirements of frontier AI development without clear paths to profitability—a challenge that, with the notable exception of the ecosystem giants, remained largely unresolved.