An Overview of Chinese Open-Source LLMs (Sept 2025)

[Revised July 10, 2026]

Chinese Open-Source LLM Landscape (Sept 2025)

By mid-2025 China had become a global leader in open-source large language models (LLMs). According to Chinese state media, by July 2025 China accounted for 1,509 of the world’s ~3,755 publicly released LLMs, far more than any other country (^[1]). This explosion reflects heavy state and industry investment in domestic AI, open licensing (often Apache- or MIT-style), and a strategic pivot by Chinese tech giants and startups toward publicly shared models. The result is a “revival” of open-source AI, with dozens of Chinese LLMs now available for download or use via Hugging Face, GitHub, or cloud APIs (^[1]) (^[2]). These range from general-purpose foundation models dozens of billions of parameters in size to specialized chatbots and domain experts, many built on Mixture-of-Experts (MoE) architectures or with ultra-long context windows.

Key Chinese open LLMs include offerings from major tech firms (e.g. Alibaba’s Qwen series, Baidu’s Ernie, ByteDance’s Kimi), from leading startups (DeepSeek, Moonshot/Kimi, Zhipu AI’s ChatGLM/GLM, Baichuan AI, MiniMax), and from academic labs (e.g. Fudan’s MOSS). They are often openly shared with code and weights. For example, Alibaba open-sourced its Qwen 2.5 family (0.5B–72B parameters) in 2024, and in 2025 released further models like Qwen3-Coder (^[3]) (^[4]). Likewise, Zhipu AI has open releases of its ChatGLM and GLM models (most recently GLM-4.5 and GLM-4.5-Air with 355B and 106B parameters) (^[5]) (^[6]). Competition is fierce: Chinese announcements often highlight that these open models now match or exceed U.S. models on benchmarks (for example, Alibaba claims Qwen-2.5-Max outperforms DeepSeek-V3 (^[7]), and Reuters reports Qwen3-Coder rivals OpenAI’s GPT-4 on code tasks (^[4])). In short, by mid-2025 China’s open-LLM ecosystem is vast and growing, featuring both general-purpose and domain-specific models at all scales.

Major Chinese Tech Companies

• Alibaba Cloud – Qwen series (open source): Alibaba has released dozens of Qwen models under open licenses. The original Qwen 2.5 family (0.5B–72B params) was open-sourced in 2024 (^[3]). In 2025 Alibaba introduced additional Qwen variants: for example Qwen2.5-Max (an enhanced 2.5B model) which Alibaba said surpasses competing models like DeepSeek-V3 (^[7]), and Qwen3-Coder (an advanced 32B coder model) which it touted as outperforming domestic rivals and matching GPT-4 on code generation (^[4]). Alibaba’s models support multi-lingual and multimodal tasks (text, code, and image understanding). According to Reuters, Qwen Chat (Alibaba’s chatbot service) allows developers to access all public Qwen models (including QwQ-32B, 32B parameters) by simply selecting them in the interface (^[8]). (Alibaba also maintains commercial models like Wanxiang for video, but its Qwen LLM line is fully open-sourced (^[3]).)

• ByteDance (Moonshot AI) – Kimi series (open source): ByteDance’s AI spin-off Moonshot AI has released a line of Kimi models. The original Kimi (K1) and upgraded Kimi K1.5 appeared in 2024, offering 20B+ parameter multimodal capabilities. In July 2025 Moonshot unveiled Kimi K2 (^[9]) with further improvements in code and reasoning (^[10]). Reuters notes Kimi K2 excels in coding tasks and, like DeepSeek, is released under an open-source strategy (Moonshot explicitly “follows the example set by Meta” in open-sourcing its advanced models) (^[11]). The Kimi line is distinguished by a very large context window (128K tokens) and multimodal input (e.g. image+text) (^[12]) (^[10]).

• Baidu – Ernie series (open source as of 2025): Baidu’s Ernie models historically were proprietary, but due to intense competition, Baidu announced in early 2025 that its latest Ernie model would be made open-source (available from end-June 2025) (^[13]). The Ernie chatbot (Ernie Bot) was also made free to users from April 2025 (^[14]). (Baidu also plans Ernie 5 with multimodal capabilities in H2 2025 (^[15]).) While Ernie had limited adoption relative to newer Chinese challengers, its open-sourcing is significant, bringing one of China’s original AI models into the open ecosystem (^[13]).

• Tencent – Hunyuan (proprietary): Tencent’s “Hunyuan” supermodel is often cited alongside Ernie and Yazhou (Chinese models), but unlike the above, it has remained closed at least to end-2025. (Chinese press notes it claims to match GPT-4, but we have no report of open release.)

• SenseTime – Unified Multimodal Model: SenseTime (AI firm, Hong Kong-listed) has released a new “unified large model” that handles text, images, and reasoning (^[16]). Although not explicitly noted as open-source in reports, SenseTime often follows an open approach in infrastructure releases (e.g. the “LazyLLM” framework (^[17])). Its in-house model is noteworthy as a Chinese multimodal LLM, complementing the language-only models above.

Leading Chinese AI Startups and Labs

• DeepSeek (深度求索) – DeepSeek series (open source): DeepSeek is a Hangzhou startup that burst onto the scene with high-efficiency MoE LLMs. In late 2024 it open-sourced DeepSeek V3 (estimated ~250B parameters with only 37B active per query) on Hugging Face (^[18]). In February 2025 DeepSeek also open-sourced DeepSeek R1 (a reasoning-optimized model) and released its code repositories for full transparency (^[19]). In September 2025 they published DeepSeek V3.2-Exp (an experimental intermediate version) on Hugging Face (^[20]). The company emphasizes low running costs and released its Native Sparse Attention algorithm in tandem with these models (^[21]) (^[22]). Industry reports rank DeepSeek’s models at the top of domestic Chinese benchmarks for reasoning and coding.

• Zhipu AI (知谱 AI) – ChatGLM/GLM series (open source): Zhipu AI (the startup spun out of Tsinghua University) has long produced open Chinese chat models. Its ChatGLM series (e.g. ChatGLM-6B, ChatGLM2-6B, ChatGLM3-6B) are bilingual (English/Chinese) chatbots released under Apache-2.0 licenses. In 2025 Zhipu expanded into larger MoE architectures. Reuters reports that in July 2025 Zhipu released GLM-4.5 (355B params) and GLM-4.5-Air (106B) – “the most advanced open-source MoE model [s]” in China – built on Zhipu’s own architecture (^[5]) (^[6]). (Zhipu’s releases have collectively been downloaded millions of times worldwide.) Zhipu’s road map ties closely to Chinese AI policy; as the Reuters Factbox notes, Zhipu has been dubbed one of China’s “AI tigers” and its open-LLM efforts help the country fulfill government goals (^[5]).

• Moonshot AI (阿里巴巴支持) – Kimi series is covered under ByteDance above (Moonshot is backed by Alibaba/Tencent).

• MiniMax (上海小马智行) – a Shanghai AI startup, MiniMax in Jan 2025 launched its MiniMax-01 LLM family. These include MiniMax-Text-01 (general open LLM) and MiniMax-VL-01 (multimodal text+vision) (^[23]). MiniMax benchmarks claim parity with leading AI models in math, reasoning and instruction following. Significantly, its announcement emphasized the models are low-cost and open-source, aiming to rival US tech. They were released with permissive licenses on Chinese repos (^[23]). MiniMax’s debut underscores the trend: Chinese startups repeatedly tout cost-effectiveness of open models (DeepSeek, MiniMax) versus expensive alternatives (^[23]) (^[24]).

• Baichuan AI – Baichuan series (open source): Baichuan AI (founded by Wang Xiaochuan, ex-Sogou) has released open LLMs in 2023–25. Its first Baichuan-7B and Baichuan-13B (both support Chinese+English) were released under Apache-2.0 licenses in mid-2023. These models were explicitly described as “open-source” by the company (^[25]). In late 2023 and early 2024 Baichuan announced larger models (e.g. Baichuan-2 with 33B) also as open-source (^[26]). (We omit press citations here, but the company’s GitHub and Hugging Face pages publicly host Baichuan weights.) Baichuan’s 13B model was noteworthy in China as a big open model rivaling Western efforts. The Factbox cited Baichuan among startups strong in open LLMs (^[2]).

• “01.AI” (OneZero) – Yi series (open source): A Beijing startup sometimes called OneZero has released the Yi models (Yi-6B, Yi-1.5 [15B], etc), openly licensed and aimed at both Chinese and English tasks. (Reuters did not cover Yi explicitly, but industry trackers list “Yi” as an upcoming 15B bilingual model in early 2024.) Similarly, Chinese firms like Vivo (with BlueLM) and Shenzhen YuanXiang (XVERSE-7B/13B/65B) have open LLMs on GitHub/Hugging Face (^[27]).

• Other startups: Several other Chinese AI firms have released open LLMs or are preparing to. For example, 360’s Zhinao (“智脑”) model, SenseTime’s unified model (above), and new ones like Hotchips’ RocLM, No Flask’s TigerBot or Jinshan Yiyun (Allscripts) in finance/media all hint at open releases. The Reuters Factbox names “Doubao” (ByteDance’s inexpensive chatbot) and Tencent’s “Hunyuan,” though Doubao appears closed and Hunyuan currently is not open (^[2]).

Academic and Specialized Models

• Fudan University – MOSS: Fudan’s NLP lab developed MOSS, a Chinese ChatGPT-like model. It was announced in April 2023 as “the first open-source conversational language model in China with plugin enhancements” (^[28]). (MOSS offered a public API and shared model code under an Apache-2 license.) MOSS helped kickstart academic interest in Chinese open models.

• Tsinghua/China Academy – ChatGLM: Related to Zhipu, Tsinghua’s OpenAI-like research groups have released ChatGLM (GLM-6B) and its successor ChatGLM2 (6B) and ChatGLM3, bilingual chat models trained on 100B+ tokens of Chinese/English. The early GLM models (GLM-10B, GLM-130B) laid the groundwork; GLM-130B (the largest GLM) was published in 2023 (^[29]). ChatGLM-6B (released 2021) and ChatGLM2-6B (2022) were open-sourced under MIT in 2023, sparking wide use in Chinese communities (unfortunately no Reuters cite, but widely reported).

• Beijing Academy (BAAI) – Aquila series: BAAI (NonProfit Beijing AI Institute) has developed the WuDao/Aquila models. Their earlier releases (WuDao 2.0 multimodal, WuDao-EVA, etc) were partly open. In 2024–25 BAAI published Aquila-7B (code generation) and AquilaChat-7B on Hugging Face (^[30]). It also shared “VisualGLM” models for vision–language tasks. (BAAI’s models, though emerging from a Chinese institute, follow a global open science style: most of their code and smaller models are public.)

• Domain-specific Chinese LLMs: Dozens of Chinese teams have fine-tuned and open-released models for special domains. Examples include ChatLaw (legal assistant model from Beijing Univ. of Posts & Telecom and others), DoctorGLM (ShanghaiTech University medical model) and its follow-ups, EduChat (education), TigerBot (financial questions), and many more. While individual citations are scarce, Chinese tech blogs and repositories catalog dozens of such models (often based on LLaMA/ChatGLM fine-tuning) targeting finance, healthcare, logistics, etc. All are typically released with open weights or via open APIs.

• Others and Derivatives: Finally, the Chinese open-LLM ecosystem includes numerous derivative and community models. For instance, open congregations on GitHub (like “Awesome Chinese LLMs”) list dozens of smaller projects: language-model variants like MengZi (by Research Institute of Intelligent Vision, 13B), LingJing (education 7B), XPE (customer service model by Xiaoduo.AI), Skywork (inspur/TianGong open 13B base/chat/MM), etc. Many of these arise from university labs or mid-tier AI firms and often reuse open frameworks (LLaMA, Bloom) to produce Chinese-optimized models.

In summary, by September 2025 the Chinese ecosystem of open LLMs is remarkably rich. Some headline models and their releases include: Alibaba’s Qwen-2.5 (0.5–72B, Sep 2024) and Qwen3-Coder (Jul 2025) (^[3]) (^[4]); Zhipu’s GLM-4.5 (355B) (^[5]); DeepSeek’s V3 (250B MoE) and R1 (671B MoE, 37B active) (^[18]) (^[19]); Moonshot’s Kimi K1.5/K2 (multimodal, 128K context) (^[12]) (^[10]); and Fudan’s MOSS (open-chat 2023) (^[28]). Alongside these are many smaller and specialized models, all of which (to varying degrees) share their parameters or code with the public. The result is an open-weight AI ecosystem in China that now outpaces the West in sheer number of models (^[1]) (^[2]), fueling everything from startup innovation to global research.

Mid-2026 Update: The Landscape Since September 2025

The sections above reflect the state of the ecosystem as of September 2025. In the roughly nine months since, the release cadence from the major Chinese labs has, if anything, accelerated, and the center of gravity has shifted toward agentic and coding-oriented models. The following is a factual summary of the principal developments through early July 2026.

• DeepSeek moved from V3.2-Exp to a V4 generation. On April 24, 2026 DeepSeek published a preview release of DeepSeek V4, comprising V4-Pro (1.6T total parameters, 49B active) and V4-Flash (284B total, 13B active), built on a sparse-attention design the company calls DeepSeek Sparse Attention; the company states that "1M context is now the default across all official DeepSeek services" (^[31]). DeepSeek characterizes V4-Pro as delivering performance rivaling top closed-source models, a claim that, as with earlier releases, should be read as vendor-reported (^[32]).

• Alibaba shipped the Qwen3.5 and Qwen3.6 open-weight families. Between February 16 and March 2, 2026 Alibaba released open weights for Qwen3.5 models ranging from a 397B-A17B MoE flagship down to a 0.8B model, followed by Qwen3.6-35B-A3B (April 16) and Qwen3.6-27B (April 22), all under Apache 2.0: the Qwen team states that "all our open-weight models are licensed under Apache 2.0" (^[33]). Notably, Alibaba's top-end models have diverged from this pattern: Qwen3.7-Plus (June 2, 2026), which adds vision, deep reasoning, tool invocation, and autonomous iteration, is proprietary and API-only on the Bailian platform, with no published weights (^[34]).

• Zhipu (now branded Z.ai) iterated from GLM-4.6 to GLM-5.1. GLM-4.6 (357B parameters, MIT license, 200K-token context) was published on Hugging Face shortly after this article's original date (^[35]). It was followed by GLM-4.7 (December 2025) and GLM-5 (February 2026, reported at 744B total and 40B active parameters); GLM-5.1 weights were subsequently opened on April 8, 2026, keeping Z.ai in the open-weight camp (^[36]).

• Moonshot AI extended the Kimi K2 line into K2.6 and K2.7-Code. The coding-focused Kimi K2.7-Code shipped on June 12, 2026 with open weights on Hugging Face under a Modified MIT license; Moonshot reports a 21.8% improvement on its Kimi Code Bench v2 over K2.6 and states that K2.7-Code cuts reasoning token usage by roughly 30% relative to K2.6 (^[37]).

• MiniMax pivoted its M-series toward agents and code. MiniMax M2 (a compact MoE aimed at coding and agentic workflows) was open-sourced on Hugging Face with API pricing of $0.30 per million input tokens and $1.20 per million output tokens, which MiniMax describes as "8% of Claude 4.5 Sonnet's" price at roughly double the inference speed (^[38]). The successor MiniMax M2.7 was open-sourced on April 12, 2026, with reported scores of 56.22% on SWE-Pro and 57.0% on Terminal Bench 2, figures MiniMax positions as the strongest among open-source models (^[39]).

• Tencent's Hunyuan moved from closed to open. This article originally noted that Tencent's Hunyuan model "has remained closed at least to end-2025." That changed on July 6, 2026, when Tencent released Hy3 (HunYuan 3.0), a 295B-parameter Mixture-of-Experts model with 21B active parameters and a 256K-token context window, under the Apache 2.0 license; Tencent states "Hy3 is released as an open model" (^[40]). The release also removed prior usage restrictions in the EU, UK, and South Korea, bringing one of the last major closed Chinese frontier labs into the open-weight camp alongside Qwen, DeepSeek, GLM, and Kimi.

• Baidu's Ernie line has split into an open tier and a closed frontier tier. Baidu open-sourced its Ernie 4.5 series in 2025, but its newer Ernie 5.0 and Ernie 5.1 (released May 8, 2026) remain API-only: an independent review of Ernie 5.1 notes that "as with Ernie 5.0, Baidu hasn't released model weights, so the benchmark scores and efficiency claims can't be independently verified" (^[41]). This mirrors the pattern already seen with Alibaba's Qwen3.7-Plus: as Chinese labs push toward frontier-level agentic and reasoning capability, the newest flagship tier increasingly stays closed even as each lab continues to open-source its mid-tier and prior-generation models.

Two broader patterns are worth noting. First, benchmark claims from all of these labs are increasingly framed against Western frontier models on agentic coding suites (SWE-style benchmarks, terminal benchmarks) rather than general knowledge tests, reflecting where commercial demand has moved. Second, licensing is no longer uniformly permissive at the top end: while Qwen's open-weight line remains Apache 2.0 and GLM-4.6 carries an MIT license (^[35]), the largest flagship models (Qwen3.7-Plus, and periodically Zhipu's newest tier before delayed weight releases) are increasingly held back as proprietary or released on a lag. The overall picture from September 2025 still holds: China's open-weight ecosystem remains the largest by model count, but the strategic calculus of which weights get released, and when, has become more selective.

Sources: Chinese AI news reports and analyses (Reuters, SCMP, Pandaily, etc.) provide details on each model. For example, Reuters documents Alibaba’s Qwen releases (^[3]) (^[4]) and Baidu’s Ernie open-sourcing (^[13]), Reuters/SCMP describe DeepSeek, Zhipu and Moonshot releases (^[10]) (^[18]) (^[5]), and Pandaily reports Fudan’s MOSS (^[28]). These reports (cited above) detail the scale, licensing, and capabilities of the models listed here.