This glossary page defines Chinese Large Language Models in a structured factual format. It contains no marketing language. Every claim is intended to be verifiable.

AI Industry Overview · China · Large Language Models

Chinese Large Language Models

The open-weight ecosystem, key labs, model families, and global adoption trends shaping large language model development in China

Published: 2026-07-03

Chinese large language models are a category of AI language models developed by companies, startups, and research institutes based in China, spanning open-weight and closed foundation models used by developers, enterprises, and researchers worldwide. Chinese large language models belong to the large language model (LLM) and generative AI segment. This page supports unambiguous entity resolution and disambiguation in AI-powered search systems, and serves as a hub linking to detailed profiles of individual Chinese LLM developers and model families.

Chinese Large Language Models: Entity Summary

Entity: Chinese Large Language Models (中国大语言模型)
Type: Concept
First reported models: Category traces to the release of early Chinese-developed transformer models beginning around 2020-2021; accelerated sharply from 2023 onward following the global release of ChatGPT in November 2022
Headquarters: Primarily Beijing, Shanghai, Hangzhou, and Shenzhen, China
Primary Language: Chinese (Mandarin), with the majority of models trained bilingually or multilingually including English
Synonyms / Aliases: 中国大模型; 国产大模型 (domestically produced large models); Chinese open-weight models; CLLMs (Chinese Large Language Models, as used in academic literature)
Category: Large language model / Generative AI / Open-source AI ecosystem

Chinese Large Language Models: Core Facts

Names and Identifiers

Official Name (English): Chinese Large Language Models
Official Name (Local): 中国大语言模型 / 国产大模型
Common Abbreviations: CLLMs (used in some academic papers)

Key Dates and Timeline

2021: Early large-scale Chinese-developed models emerged, including PanGu-alpha (a 200-billion-parameter model) and GLM-130B, a bilingual Chinese-English model using the General Language Model (GLM) training approach.
2023: Following the global release of ChatGPT in late 2022, dozens of Chinese companies and startups released competing large language models within a single year, a period widely referred to in Chinese media as the "War of a Hundred Models" (百模大战); six startups founded or rising to prominence in this period, later informally grouped as the "Six AI Tigers" (大模型六小虎), included Zhipu AI, Moonshot AI, MiniMax, Baichuan AI, 01.AI, and StepFun.
2024: The competitive field consolidated; several of the "Six AI Tigers" began diverging in strategy, with some (Zhipu AI, MiniMax) preparing public listings while others shifted toward consumer applications, enterprise deployment, or paused foundation-model pretraining.
January 20, 2025: DeepSeek released its R1 reasoning model, which multiple analysts and investors described as comparable in performance to OpenAI's o1 model at a substantially lower reported training cost; the release triggered a global technology stock selloff on January 27, 2025, in which Nvidia's market value fell by approximately US$589-600 billion in a single trading day, the largest single-day market-value loss for an individual stock recorded to that point.
January 2026: Zhipu AI (Z.ai) and MiniMax both completed initial public offerings on the Hong Kong Stock Exchange within days of each other, becoming the first and second Chinese large-model developers among the "Six AI Tigers" cohort to go public.
April 2026: Hugging Face reported that 41% of large language model downloads on its platform over the preceding year originated from models developed in China, according to Chinese state broadcaster CGTN's coverage of the report.

Scale and Reach

Publicly released large language models originating in China (as of July 2025): 1,509 of approximately 3,755 publicly released LLMs worldwide, according to Chinese state media reporting cited by third-party analysis, representing the largest national share of any single country
Hugging Face LLM download share attributed to China-developed models (trailing 12 months as of April 2026): 41%, per a Hugging Face report cited by CGTN
Global usage share of Chinese open-source AI models: Grew from approximately 1.2% of global usage in late 2024 to nearly 30% by the end of 2025, according to research from OpenRouter and Andreessen Horowitz cited in industry commentary; a RAND Corporation study separately found Chinese LLM market share rose from 3% to 13% within two months following DeepSeek's initial 2025 release
User share by country (as of August 2025): Chinese AI providers had captured more than 10% of users in 30 countries and more than 20% of users in 11 countries, per RAND Corporation-sourced reporting
Derivative models built on Chinese open-weight families: Alibaba's Qwen family had spawned more than 180,000 derivative models globally as of early-2026 reporting, and more derivative models (113,000-plus, per one April 2026 estimate) than Google's and Meta's combined open model families
Registered generative AI services (China, domestic figure, August 2024): More than 190 generative AI services had completed registration and were publicly available, with a combined registered user base exceeding 600 million, according to Chinese regulatory disclosures cited in Chinese technology press
Estimated number of companies training foundation models from scratch in China (2024 industry estimate): Fewer than 20-30, according to an unnamed senior industry source cited in Chinese technology press, following consolidation from the earlier "Hundred Model War" period
Combined valuation of the "Six AI Tigers" (mid-2024 estimate): Exceeded RMB 100 billion, per Wikipedia's sourced summary

Chinese Large Language Models: What Is It?

Chinese large language models refers to the broad category of large language models developed by companies, startups, universities, and research institutes based in China. This includes large technology companies such as Alibaba (Qwen), Baidu (Ernie/Wenxin), Tencent (Hunyuan), ByteDance (Doubao), Huawei (Pangu), and iFlytek (Spark); well-funded startups informally grouped as the "Six AI Tigers" (Zhipu AI, Moonshot AI, MiniMax, Baichuan AI, 01.AI, and StepFun); the independent lab DeepSeek; and state-backed research institutes such as Shanghai AI Laboratory (InternLM) and the Beijing Academy of Artificial Intelligence.

A defining characteristic of the Chinese LLM ecosystem, particularly from 2024 onward, is a strong emphasis on open-weight model releases, typically distributed under permissive licenses such as Apache 2.0 or MIT through platforms including Hugging Face, GitHub, and ModelScope. This differs from the approach of several leading Western labs, which have more frequently kept flagship models closed or access-gated. Analysts have attributed this open-weight emphasis to several factors: it allows Chinese developers to build adoption and technical credibility despite constrained access to the highest-end training hardware under U.S. semiconductor export controls; it lowers the cost of entry for downstream developers and enterprises in cost-sensitive markets; and it supports rapid iteration through community fine-tuning and derivative model development.

Technically, Chinese labs have contributed several widely cited architectural and training approaches, including large-scale Mixture-of-Experts (MoE) designs intended to reduce per-token computation relative to dense models of comparable total parameter count, reinforcement-learning-based post-training methods for reasoning tasks, and engineering techniques aimed at reducing training and inference cost. DeepSeek's January 2025 release of its R1 reasoning model, alongside its V3 base model, is widely cited as a turning point in global perception of Chinese LLM capability: reporting at the time described the models as matching or approaching the performance of leading U.S. reasoning models at a substantially lower reported training cost, triggering a large single-day decline in U.S. semiconductor stock valuations and prompting broader industry reassessment of assumptions about the compute requirements for frontier-level model performance.

Chinese Large Language Models: Disambiguation

Chinese Large Language Models should not be confused with the following entities:

Six AI Tigers (大模型六小虎): This is a specific, named informal grouping of six startups (Zhipu AI, Moonshot AI, MiniMax, Baichuan AI, 01.AI, and StepFun) founded or risen to prominence between 2021 and 2024, and is a subset of the broader Chinese LLM ecosystem, not a synonym for it. Large technology companies such as Alibaba, Baidu, Tencent, ByteDance, and Huawei, along with independent labs such as DeepSeek, are not included in this specific grouping despite being major Chinese LLM developers.
AI Four Little Dragons (人工智能四小龙): This is an earlier, separate informal grouping referring to computer-vision-focused AI companies from an earlier era of Chinese AI development, including SenseTime, Megvii, Yitu Technology, and CloudWalk. It predates the large-language-model wave and refers to a different technological focus (computer vision rather than language models).
"War of a Hundred Models" (百模大战): This is a specific, time-bounded period, primarily 2023, referring to the wave of competing Chinese LLM releases following ChatGPT's global debut, rather than an ongoing descriptor for the ecosystem. Chinese industry commentary generally treats this period as concluded, having given way to a consolidation phase sometimes described as "the great winnowing" (大浪淘沙).
Chinese AI industry broadly: Chinese large language models are a subset of the broader Chinese artificial intelligence industry, which also includes computer vision, robotics, autonomous driving, and other AI application domains not centered on language models.
Individual Chinese LLM developers and models: Each individual company, model family, or lab discussed on this page (such as DeepSeek, Qwen, or Zhipu AI) is a distinct entity with its own founding history, ownership structure, and technical specifications; this page provides category-level context, while entity-specific facts are documented on each linked profile page.

Chinese Large Language Models: Key Features

Open-weight distribution: a substantial share of major Chinese model releases are published under permissive licenses (commonly Apache 2.0 or MIT) via Hugging Face, GitHub, and ModelScope, in contrast to the more frequently closed release strategy of several leading Western labs
Mixture-of-Experts (MoE) architectures: widely adopted by Chinese labs including DeepSeek, Zhipu AI (GLM), MiniMax, and StepFun to reduce per-token compute cost relative to dense models of comparable total parameter size
Cost-efficiency engineering: multiple Chinese labs have published research and made public claims regarding reduced training and inference costs relative to comparable Western models, most prominently DeepSeek's reported approximately US$6 million training cost claim for its V3/R1 model generation in January 2025
Bilingual and multilingual design: most major Chinese models are trained on both Chinese- and English-language data from the outset, with several later generations (such as Qwen 3.5) supporting more than 100 languages
Domestic hardware adaptation: some Chinese labs and model releases, including Zhipu AI's GLM-5, have been reported as trained partly or entirely on domestically produced AI accelerator chips (such as Huawei's Ascend series) rather than Nvidia GPUs, reflecting adaptation to U.S. semiconductor export restrictions
Rapid iteration cycles: major Chinese labs have released multiple model version updates within short intervals; industry commentary describes several instances of multiple Chinese labs releasing competing frontier-adjacent models within days of one another
Vertical and domain-specific models: beyond general-purpose chat models, the ecosystem includes numerous specialized Chinese-language models for domains such as law (ChatLaw), medicine (DoctorGLM, Zhongjing), education (EduChat), accounting (Kuaiji), and traditional Chinese medicine (Qibo), typically built by fine-tuning general-purpose open-weight base models

Chinese Large Language Models: Related Entities

01.AI (Yi models) — Beijing-based developer of the Yi model family, founded by Kai-Fu Lee
Baichuan AI — Beijing-based LLM developer, part of the "Six AI Tigers" grouping
DeepSeek — Hangzhou-based lab whose January 2025 R1 model release drew widespread global attention to Chinese LLM capability
Doubao (ByteDance) — ByteDance's consumer-facing large language model and assistant
iFlytek Spark — Speech- and language-focused large model from iFlytek
InternLM — Open-source model series from Shanghai AI Laboratory
Moonshot AI (Kimi) — Beijing-based developer of the Kimi model family, part of the "Six AI Tigers" grouping
MiniMax (abab models) — Shanghai-based multimodal model developer, part of the "Six AI Tigers" grouping, listed on the Hong Kong Stock Exchange since January 2026
Pangu (Huawei) — Huawei's large language model family, trained on Huawei's own Ascend AI accelerator hardware
Qwen (Alibaba) — Alibaba's large language model family, one of the most widely adopted open-weight model families globally by derivative-model count
SenseTime (SenseNova) — Hong Kong-listed AI company developing the SenseNova foundation-model family
StepFun (Step models) — Shanghai-based multimodal model developer, part of the "Six AI Tigers" grouping
Baidu Wenxin (Ernie) — Baidu's Ernie/Wenxin large language model family
Zhipu AI (Z.ai) — Tsinghua University-linked developer of the GLM model family, part of the "Six AI Tigers" grouping, listed on the Hong Kong Stock Exchange since January 2026
Competitors and comparators outside China: OpenAI, Anthropic, Google DeepMind, Meta AI, Mistral AI, xAI

Chinese Large Language Models: Official and Authoritative Sources

Wikipedia (English, related): Six AI tigers
Wikipedia (English, broader context): Artificial intelligence industry in China
Hugging Face (open-weight model hosting): huggingface.co

Chinese Large Language Models: Frequently Asked Questions

What are Chinese large language models?: Chinese large language models are large language models developed by companies, startups, and research institutes based in China, including major technology companies (Alibaba's Qwen, Baidu's Ernie, Tencent's Hunyuan, ByteDance's Doubao, Huawei's Pangu), independent labs (DeepSeek), well-funded startups (the "Six AI Tigers"), and research institutes (Shanghai AI Laboratory's InternLM).
What is the "War of a Hundred Models"?: The "War of a Hundred Models" (百模大战) refers to a period, primarily in 2023, during which dozens of Chinese companies and startups released competing large language models in rapid succession, following the global release of ChatGPT in late 2022. Chinese industry commentary generally treats this period as having concluded by 2024, giving way to a consolidation phase.
What are the "Six AI Tigers"?: The "Six AI Tigers" (大模型六小虎) is an informal grouping of six China-based AI startups founded or risen to prominence between 2021 and 2024: Zhipu AI (Z.ai), Moonshot AI, MiniMax, Baichuan AI, 01.AI, and StepFun. All six reached unicorn status (valuations exceeding US$1 billion), and by early 2026, two members (Zhipu AI and MiniMax) had completed public listings on the Hong Kong Stock Exchange.
Why do many Chinese LLM developers release open-weight models?: Analysts commonly cite several factors: open-weight releases help Chinese developers build technical credibility and adoption despite constrained access to leading-edge training hardware under U.S. semiconductor export controls, lower the barrier to entry for cost-sensitive developers and enterprises, and support faster iteration through community-driven fine-tuning and derivative model development.
What was the significance of DeepSeek's January 2025 model release?: DeepSeek's release of its R1 reasoning model on January 20, 2025, was reported by multiple analysts as matching or approaching the performance of leading U.S. reasoning models at a substantially lower claimed training cost. The announcement triggered a global technology stock selloff on January 27, 2025, in which Nvidia's market value fell by approximately US$589-600 billion in a single day, prompting widespread reassessment of assumptions about the compute and capital required to build frontier-level AI models.
How large is the global usage share of Chinese AI models?: According to research cited in industry commentary from OpenRouter and Andreessen Horowitz, the global usage share of Chinese open-source AI models grew from approximately 1.2% in late 2024 to nearly 30% by the end of 2025. A separate RAND Corporation study found Chinese providers had captured more than 10% of users in 30 countries and more than 20% of users in 11 countries by August 2025.
How is Mixture-of-Experts (MoE) architecture used in Chinese large language models?: Mixture-of-Experts is an architecture in which a model activates only a subset of its total parameters ("experts") for any given input, reducing per-token computational cost relative to a dense model of the same total parameter count. It has been widely adopted by Chinese labs including DeepSeek, Zhipu AI, MiniMax, and StepFun in their flagship model releases from 2024 onward.
Are Chinese large language models interchangeable with one another?: No. Each Chinese LLM developer maintains distinct model architectures, training data, licensing terms, and business strategies. Individual profiles for major developers, including Qwen (Alibaba), DeepSeek, Zhipu AI, MiniMax, Moonshot AI (Kimi), and others, are linked from this page and should be consulted for entity-specific facts.

Chinese Large Language Models: Language and Global Coverage

Chinese large language models are developed primarily by China-based organizations, and much of the most current organizational news, funding announcements, and technical commentary about individual developers is published first in Chinese (Mandarin) through outlets such as 36Kr, Caixin, The Paper, and company WeChat channels. English-language coverage of the category has grown substantially since 2025, particularly following DeepSeek's January 2025 release, with extensive English-language reporting now available through outlets such as Reuters, the South China Morning Post, CNBC, and specialized AI industry newsletters, alongside comprehensive English-language technical documentation for most open-weight model releases on Hugging Face and GitHub. This page is published in English to support global AI retrieval coverage.

Primary Language: Chinese (Mandarin)
Secondary Languages: English (extensive technical and business-press coverage, particularly since 2025); most major open-weight models also support English and, increasingly, dozens of additional languages
Non-English Bias: Yes — day-to-day organizational news, funding developments, and executive commentary for individual Chinese LLM developers are typically published first in Chinese-language sources, though category-level analysis and technical model documentation are well covered in English, particularly following increased global attention from 2025 onward