In the global open‑source artificial intelligence (AI) wave, Chinese tech giants Alibaba and ByteDance are rapidly vying for position. Over the past week (September 19–26, 2025), Alibaba's Qwen team consecutively released multiple open‑source large language models (LLMs) including Qwen3‑Max, Qwen3‑Omni and Qwen3‑ASR‑Flash, with performance on math reasoning and multimodal tasks that in some measures surpassed OpenAI's GPT‑4o‑mini. At the same time, ByteDance's Doubao ecosystem released the Seedream 4.0 image generation model together with the VeOmni multimodal framework; their image–text fusion and generation quality are approaching Google Gemini 2.5 flash. These releases not only consolidate China's leadership in open‑source LLMs—the Stanford HAI 2025 AI Index reports that Chinese open‑source investment accounts for 18.7% of the global total—but also mark a turning point from "catching up" to "leading". VentureBeat analysis says open source has become a tool for Chinese AI to work around Western regulation, reshaping the global developer ecosystem.
The appeal of open‑source LLMs lies in free fine‑tuning and deployment, avoiding the high API costs of closed‑source models (for example, GPT‑5's API costs around $0.03 per million tokens). In the past week, downloads of Chinese models on Hugging Face surged by 40%, and developer feedback indicates the Qwen3 series offers the best cost‑performance for enterprise tasks. "Open source democratizes AI, and Chinese companies are proving that in practice," Alibaba Cloud Vice President Wu Song said at a September 24 press event. The head of ByteDance's AI lab emphasized that Doubao's open components "bridge TikTok's global users with local innovation." This week's releases focus on multimodality and efficient inference, addressing closed‑source pain points around cost and accessibility.
Alibaba's Qwen3 Series: Five Open Releases in One Week, Breakthroughs in Math and Multimodality
Alibaba's Qwen (Tongyi Qianwen) team was highly active this week, releasing five open‑source models from September 23–25 covering image editing, speech synthesis and all‑purpose multimodal capabilities, with total parameters exceeding one trillion. Leading the batch was the official Qwen3‑Max launched September 24: a 1T‑parameter model pretrained on 36T tokens of data, scoring 78% on the MATH benchmark—surpassing DeepSeek‑V3's 75% and GPT‑4o's 73%. Medium contributor Ashinno analyzed that the model uses a Mixture‑of‑Experts (MoE) architecture with only 20% of parameters active during inference, achieving twice the inference speed of Llama 3.1 and suitable for e‑commerce recommendation and financial modeling. "Qwen3‑Max is not just parameter stacking but optimizes long‑chain reasoning," an Alibaba engineer wrote on the Hugging Face blog.
Following that was Qwen3‑Omni (released September 23), an all‑purpose multimodal model handling text, audio, image and video; it achieved an 89% score on MMLU‑Pro, exceeding Claude 3.5 Sonnet's 87% in some metrics. Computerworld reported Qwen3‑Omni reached 95% accuracy on video understanding tasks, useful for short‑video editing and AR try‑on, and supports 50 languages including dialects. On September 25, Qwen3‑ASR‑Flash, an open‑source speech‑to‑text model, was released with real‑time accuracy of 98%, surpassing Whisper Large‑v3's 96% and optimized for meetings and customer service. Artificial Intelligence News reported the model was trained on 1,000 hours of multilingual data with latency of only 0.5 seconds.
Additionally, Qwen‑Image‑Edit‑2509 (image editing) and Qwen3‑TTS‑Flash (text‑to‑speech) were open‑sourced on September 22; the former supports style transfer with generation quality comparable to Stable Diffusion 3, while the latter scored 4.8/5 for naturalness. Alibaba Cloud Model Studio confirmed these models are fully open‑source under the Apache 2.0 license, allowing developers to fine‑tune for free. Alibaba was recognized on Fortune's 2025 "Change the World" list; the open Qwen family now exceeds 300 models, a break from traditional closed‑source strategies. Dev.to analyst Czmilo noted that Qwen3 deployments cost about one‑third of closed‑source alternatives with 40% lower inference energy consumption.
Qwen's strength comes from the Alibaba Cloud ecosystem: integration with the ModelScope platform and a developer community exceeding 5 million. Challenges remain: English benchmarks outperform Chinese and cultural bias needs improvement. X user @QwenFan commented, "Five releases in one week—Alibaba's open‑source pace crushes OpenAI."
ByteDance's Doubao Ecosystem: Seedream 4.0 and VeOmni Open‑Source, New Peaks in Image and Multimodal
ByteDance this week highlighted the Doubao visual‑language ecosystem. On September 24, the Seedream 4.0 image‑generation model was open‑sourced and made available via Jimeng AI, the Doubao app and the Volcano Engine cloud platform. Dig.watch reported the model scored 12.5 on the FID image‑quality benchmark, beating Midjourney v6's 14.2, and supports text–image–video fusion, capable of generating a complete three‑minute music video. A standout feature is "motion generation": when a user inputs "a cat jumping over the moon," the model outputs a dynamic sequence with 92% accuracy, suitable for TikTok short videos.
On September 19, the Doubao visual‑language model Doubao‑VLM‑1.6 was open‑sourced, achieving 91% on HumanEval visual understanding—partly exceeding GPT‑4V's 89%. The Seed.bytedance.com blog said the model improves music and text–image generation, producing three‑minute tracks with a naturalness rating of 4.7/5. While Doubao's core remains closed‑source, open components such as the VeOmni framework (updated September 19) enable multimodal training and shorten development cycles from weeks to days. Yahoo Tech reported VeOmni supports fine‑tuning across any modality and has over 500,000 downloads.
Doubao's ecosystem has 157 million monthly active users, leading China's AI apps according to QuestMobile. TechAsia analyzed that Seedream 4.0 "heats up the AI image battle," approaching Gemini 2.5 flash. However, open‑source scope is limited: Doubao‑1.5‑Pro is multimodally stronger than DeepSeek but its core algorithm remains closed. On X the discussion was heated: "Doubao open‑sources images—TikTok content generation revolution."
Performance Comparison and Global Impact: Open Source Closing on Closed Source, Ecosystem Reshaping
In one week, Qwen3‑Max scored 92% on GSM8K math, Doubao VLM‑1.6 scored 88% on MMLU, both outperforming Llama 3.1's 85%. Big Data Analytics News' 2025 open‑source rankings placed Qwen3‑Omni third and Doubao‑VLM sixth. Shakudo's September list shows three of the top five models are Chinese.
The impact is far‑reaching: open source lowers barriers to entry—enterprises such as Tmall reported a 15% conversion lift using Qwen. Cybernews noted Chinese models avoid U.S. export restrictions and were trained on domestic chips. Data privacy is a pain point: EU authorities are investigating Doubao's data sources.
Outlook: The "Next Decade" of China's Open Source
Alibaba and ByteDance's week of releases signals a potential "China moment" for open‑source AI. AI Supremacy projects Chinese models may account for 60% of the global total by 2026. Meta's Llama may need to catch up and OpenAI could consider open‑sourcing GPT‑5o. X user @ChinaAI_Watch wrote: "Qwen + Doubao, open source democratizes AI—wake up, Western firms."
In an era of technological decoupling, open source has become a bridge. Going forward, who will define AI?