The latest list of the top 16 Chinese AI open-source players is out.
It's getting too competitive! It's getting too competitive!
The results of the well - known AI large model evaluation, Chatbot Arena, are out!
Alibaba's Qwen3 - 235B - A22B - Instruct - 2507 ranks third on the overall list of large language models. Dark Side of the Moon's Kimi - K2 - 0711 - preview and DeepSeek's DeepSeek - R1 - 0528 are tied for fifth on the overall list, surpassing top - tier closed - source models such as Claude 4 and GPT - 4.1 in an open - source manner.
Among multi - modal large models, ByteDance's image generation model Seedream - 3 ranks fifth on the text - to - image list. ByteDance's image editing model Seededit - 3.0, ByteDance's open - source unified multi - modal foundation model BAGEL, and StepStar's open - source image editing model Step1X - Edit rank fifth, seventh, and eighth respectively on the image editing list.
In the Chatbot Arena programming list announced on July 30th, DeepSeek's Deepseek - R1 - 0528 and Alibaba's Qwen3 - Coder are tied for first, and Dark Side of the Moon's Kimi - K2 - Instruct ranks seventh.
Among them, Alibaba's Qwen3 - 235B - A22B - Instruct - 2507, Alibaba's Qwen3 - Coder, and Dark Side of the Moon's Kimi K2 are all open - source models newly released in July.
At the end of July and the beginning of August, Chinese AI large models are still being updated intensively. ByteDance, StepStar, Alibaba, and Dark Side of the Moon have successively announced new model actions in the past two days.
On Thursday night, ByteDance's Seed team released the experimental diffusion language model Seed Diffusion Preview, and StepStar open - sourced the foundation large model Step3. On Friday, Alibaba open - sourced the "dessert - level" programming model Qwen3 - Coder - 30B - A3B - Instruct, and Dark Side of the Moon released the high - speed version of kimi - k2, kimi - k2 - turbo - preview.
Hugging Face also released the highlights of the Chinese AI community in July on Friday, looking back on the dazzling open - source wave in this month, including:
✨ Another "DeepSeek Moment" — Kimi K2
✨ Qwen's full matrix - Instruct / Thinking / Coder models span a parameter scale from 30B to 480B
✨ The multi - modal wave:
- GLM - 4.1V - Thinking: Image+Text > Text
- Intern - S1: Image+Text > Text
- Wan 2.2 - Text +Image > video
- Skywork - R1V3: Image+Text > Text
- Skywork - UniPic: Text > Image / Image > Text
- Tar - 7B: Any - to - Any
- Ming - Lite - Omni - 1.5: Any - to - Any
- Step3: Image+Text > Text
- HunyuanWorld - 1: Image > 3D
- ThinkSound: Video > Audio
- Neta - Lumina: Text > Image
✨ Lightweight and deployable models
- SmallThinker runs on 1GB RAM
✨ Agentic programming becomes mainstream
- Qwen3 - Coder: fully spec'd tool calling
- GLM - 4.5: browser agents, IDE assistant
- Qwen3 WebDev demo: text - to - frontend code
✨ Domain - specific and practical models/tools/datasets
- Science one S1: Scientific model
- Agentar DeepFinance: Finance dataset
- ObjectClear: Interactive Vision Tool
- Qwen3 MT Demo: Machine Translation Tool
Among them, the 31 eye - catching open - source models, 1 framework, and 1 dataset reviewed in July come from 16 enterprises, universities, or research institutions: Alibaba (9), Dark Side of the Moon (2), Zhipu (2), StepStar (1), ByteDance (2), Kunlun Tech (2), ZhiYuan Research Institute (1), China Telecom AI Research Institute (1), Ant Group (4), Kuaishou (1), Neta (1), Panshi (3), Shanghai Jiao Tong University (1), Tencent (1), Shanghai AI Laboratory (1), Fudan University (1).
1. Alibaba (9): Programming models Qwen3 - Coder - 30B - A3B - Instruct, Qwen3 - Coder - 480B - A35B - Instruct; deep - thinking models Qwen3 - 30B - A3B - Thinking - 2507, Qwen3 - 235B - A22B - Thinking - 2507; foundation models Qwen3 - 235B - A22B - Instruct - 2507, Qwen3 - 30B - A3B - Instruct - 2507; CoT audio model ThinkSound; unified video generation model Wan2.2 - TI2V - 5B; text - to - video model Wan2.2 - T2V - A14B.
2. Dark Side of the Moon (2): MoE foundation model Kimi - K2 - Base; a mathematical theorem - proving model Kimina - Prover - 72B jointly developed with the Numina team.
3. Zhipu (2): Multi - modal large model GLM - 4.1V - 9B - Thinking; foundation model GLM - 4.5.
4. StepStar (1): Foundation model Step3.
5. ByteDance (2): Agent model Tar - 7B; multi - language translation model Seed - X - Instruct - 7B.
6. Kunlun Tech (2): Multi - modal reasoning large model Skywork - R1V3 - 38B; multi - modal unified model Skywork - UniPic - 1.5B.
7. ZhiYuan Research Institute (1): Text - to - voice - over video framework MTVCraft.
8. China Telecom AI Research Institute (1): AI - Flow - Ruyi - 7B - Preview0704.
9. Ant Group (4): Multi - modal reasoning model M2 - Reasoning; multi - modal large model Ming - Lite - Omni - 1.5; financial training dataset Agentar - DeepFinance - 100K; interactive deep - reasoning model KAG - Thinker - en - ch - 7b - instruct.
10. Kuaishou (1): Adaptive thinking model KAT - V1 - 40B.
11. Neta (1): Anime - style image generation model Neta - Lumina.
12. Panshi (3): Scientific foundation large models S1 - Base - 671B, S1 - Base - 8B, S1 - Base - 32B.
13. Shanghai Jiao Tong University (1): Edge - native large model SmallThinker - 4BA0.6B - Instruct.
14. Tencent (1): 3D world generation model HunyuanWorld - 1.
15. Shanghai AI Laboratory (1): Scientific multi - modal large model Intern - S1.
16. Fudan University (1): Voice generation model MOSS - TTSD - v0.5.
Two days ago, the top ten on the Hugging Face trend list were completely occupied by Chinese - origin teams. In the latest list, Chinese - origin teams occupy 8 of the top ten spots. Among them, Zhipu takes 2 spots, Tencent takes 1 spot, Alibaba takes 4 spots, and BosonAI, an American startup founded by AI expert Li Mu, takes 1 spot.
From the latest Hugging Face trend list TOP100, we sorted out the list of Chinese AI open - source teams, mainly including 14 enterprises and independent research institutions, namely Zhipu, Tencent, Alibaba, Shanghai AI Laboratory, Dark Side of the Moon, StepStar, Neta, Kunlun Tech, Xiaohongshu, ByteDance, Ant Group, DeepSeek, Kuaishou, ZhiYuan Research Institute, as well as universities such as Shanghai Jiao Tong University (SmallThinker), Fudan University (co - operating OmniSVG with StepStar), City University of Hong Kong (PusaV1).
The models of enterprises such as HiDream, Meituan (MeiGen), SketchUp (SpatialLM), Baidu (ERNIE), MiniMax, Panshi (ScienceOne), and Mianbi Intelligence (MiniCPM) are outside the top 100 of the latest trend list, but they are also important forces in the Chinese open - source AI ecosystem.
In total, there are more than 20 Chinese enterprises actively embracing AI open - source.
Especially Alibaba is the undisputed number - one "competitive player". The frequency of its competitiveness is far ahead. In July, it successively open - sourced the audio model ThinkSound, the multi - modal reasoning model HumanOmniV2, the programming model Qwen3 - Coder (480B - A35B - Instruct), the reasoning models Qwen3 - 235B - A22B - Thinking - 2507, Qwen3 - 30B - A3B - Thinking - 2507, the foundation model Qwen3 - 235B - A22B - Instruct - 2507 (non - thinking version), the movie - level video model Wan2.2, the programming model Qwen3 - Coder - Flash (30B - A3B - Instruct), and made the new GSPO algorithm behind the new model public.
The following are the Chinese enterprises, independent research institutions, and the models on the latest Hugging Face trend list TOP100:
1. Zhipu (7): GLM - 4.5, GLM - 4.5 - Air, GLM - 4.5 - FP8, GLM - 4.5 - Base, GLM - 4.5 - Air - FP8, GLM - 4.1V - 9B - Thinking, GLM - 4.5 - Air - Base.
2. Tencent (3): HunyuanWorld - 1, Hunyuan3D - 2.1, X - Omni - En.
3. Alibaba (18): Qwen3 - 30B - A3B - Instruct - 2507, Qwen3 - Coder - 480B - A35B - Instruct, Qwen3 - Coder - 30B - A3B - Instruct, Qwen3 - 235B - A22B - Thinking - 2507, Wan2.2 - TI2V - 5B, Wan2.2 - T2V - A14B, Qwen3 - 30B - A3B - Thinking - 2507, Wan2.2 - I2V - A14B, Qwen3 - 235B - A22B - Instruct - 2507, Wan2.2 - TI2V - 5B - Diffusers, Owen3 - 30B - A3B - Instruct - 2507 - FP8, Wan2.2 - T2V - A14B - Diffusers, Wan2.2 - I2V - A14B - Diffusers, Qwen3 - Coder - 480B - A35B - Instruct - FP8, Qwen3 - 0.6B, Qwen3 - 235B - A22B - Thinking - 2507 - FP8, Qwen3 - 30B - A3B, Qwen3 - 8B.
4. Shanghai AI Laboratory (2): Intern - S1, Intern - S1 - FP8.
5. Dark Side of the Moon (1): Kimi - K2 - Instruct.
6. StepStar (2): Step3, OmniSVG (co - operated with Fudan University).
7. Neta (1): Neta - Lumina.
8. Kunlun Tech (1): Skywork - UniPic - 1.5B.
9. Xiaohongshu (1): dots.ocr.
10. ByteDance (1): Seed - X - PPO - 7B.
11. Ant Group (1): Ming - Lite - Omni - 1.5.
12. DeepSeek (2): DeepSeek - R1, DeepSeek - R1 - 0528 - Qwen3 - 8B.
13. Kuaishou (1): KAT - V1 - 40B.
14. ZhiYuan Research Institute (1): bge - m3.
Some well - known individual developers also