Alibaba's open source Qwen 3 redefines global access with 119 languages, beating ChatGPT and Gemini
The world has already witnessed China's "DeepSeek moment," a significant marker of its rapidly advancing artificial intelligence (AI) capabilities, and the wave of innovation continues to surge.
In late April, Chinese tech giant Alibaba dropped its newest large language model, Qwen 3, which has been turning heads, especially among developers. It’s packed with interesting features and, more importantly, the model is open-source under the Apache 2.0 license, meaning anyone can use, tweak, or integrate it freely.
Chinese companies are eyeing the global stage, and Qwen 3’s strong multilingual capabilities and developer-friendly design show Alibaba is serious about keeping pace with American heavyweights like OpenAI and Google in the AI arena.
What makes Qwen 3 different?
In a blog post, Qwen 3 highlights one of its key features, which is a “hybrid thinking” approach to answering questions. Basically, the AI has two modes: “thinking mode” for tougher problems that need more brain power and longer delivery time, and “non-thinking mode” for quick, straight-to-the-point answers. It lets users pick between depth and speed, depending on the task. This level of user control over reasoning depth is still rare among top AI models.
Another feature that stands out is that Qwen 3 supports a whopping 119 languages and dialects – not just English, Mandarin, French and other languages used in the world of big business, but also local languages from Southeast Asia like Javanese and Minangkabau in Indonesia, as well as Cebuano and Waray in the Philippines. Even lesser-used languages like Haitian Creole and Ta’izzi-Adeni Arabic make the list.
That kind of range makes Qwen 3 especially handy for developers building tools for users across Asia, Africa, and other regions who may not be fluent in international languages.
By comparison, ChatGPT supports nearly 60 languages while Google’s Gemini supports about 40, with the quality varying significantly depending on the lingo.
The performance of large language models or LLMs is generally strongest in high-resource languages like English and Mandarin. For other languages, the results may not be as accurate.
In the case of Javanese, used by over 90 million people in Indonesia, ChatGPT and Qwen handle basic Javanese language prompts fairly well but both fall short when it comes to culturally embedded expressions.
Unlike ChatGPT, Qwen 3 lists Javanese among its supported languages, indicating that the model has been intentionally trained and tuned for it. While this doesn’t guarantee fluency, it suggests stronger potential for improvement over time.
The inclusion of many more local and low-resource languages in Qwen 3 reflects Alibaba’s efforts to reach out to a broader audience compared to its US rivals.
Finally, Qwen 3 is built using something called a Mixture-of-Experts [MoE] design which helps it run more efficiently. MoE works by breaking a task into parts and assigning each one to a smaller, specialized “expert” model.
Not perfect yet
Qwen 3 has its drawbacks. For instance, its larger models require high-end hardware while its “thinking mode” can slow things down. The cost of running Qwen 3 is a lot lower, however.
According to reviews compiled by Substack writer and tech entrepreneur Patrick McGuinness, one of the biggest complaints about Qwen 3 is its tendency to overthink. Some programmers say the models tend to “think too much,” making them a bit slower and less reliable for tasks like coding.
In a YouTube review, AICodeKing pointed out that Qwen 3 can get stuck in loops. While he sees potential in the model, he believes it still falls short of DeepSeek R1 when it comes to coding tasks.
In a global ranking of AI models called LiveBench, Qwen 3 competes closely with top models. It was slightly behind OpenAI’s o3, Google’s Gemini Pro 2.5, and Anthropic’s Claude 3.7, but more advanced than DeepSeek R1 and Grok 3 Mini Beta (High).
OpenAI’s o3-mini high came out on top as the best overall model.
But when it comes to cost, Qwen 3 is much cheaper to use. Running o3 costs about $10 per 1 million tokens, while Qwen 3 costs only 4 yuan (about $0.55) for the same amount, according to the South China Morning Post. This makes Qwen 3 a more budget-friendly option.
While there’s room for improvement, Qwen 3 shows plenty of promise.
Why the world should pay attention
With such capabilities, Alibaba’s Qwen 3 isn’t just another AI model out of China; it’s a serious contender with performance that rivals big names such as OpenAI and DeepSeek.
Qwen 3’s open-source release is another sign of how important open access has become in China’s AI push. Earlier this year, DeepSeek also drew global attention with its open model, giving developers room to explore different use cases.
In contrast, top American LLMs like Google’s Gemini, Anthropic’s Claude, and OpenAI’s GPT-4 remain closed. In February, former Google CEO Eric Schmidt warned that the West must invest in open-source AI or risk falling behind China. Open access encourages broader contributions, accelerating progress in the AI industry.
Making AI tools open also lowers the barrier, especially for smaller companies and researchers in emerging markets like Southeast Asia’s to build, adapt, and experiment with models for local needs.
In a LinkedIn post, Pavitra Mukherjee, country head at eduCLaaS Vietnam, said open-source AI is crucial for Southeast Asia. It cuts out costly licensing fees, supports local languages and culture, speeds up localization, and reduces reliance on Western tech.
Southeast Asia has seen several key open-source AI projects. Singapore’s SEA-LION is a family of LLMs built to understand the region’s diverse languages and cultures. Vietnam’s VinAI is also contributing open-source tools, while in Indonesia, Indosat Ooredoo Hutchison and GoTo have partnered to develop their own open-source model, Sahabat AI.
With the arrival of advanced tools like Qwen 3, developers across the region now have even more powerful, flexible resources to build with.
In countries where many local languages are spoken, Qwen 3’s 119 languages offer a distinct advantage for developers that want to serve the entire population.
In their paper “Speaking in Code: Contextualizing Large Language Models in Southeast Asia,” Carnegie Endowment’s Asia fellows Elina Noor and Binya Kanitroj explain that language in the region is closely tied to identity, culture, and politics.
Many Southeast Asian languages are considered “low-resource” due to the limited amount of information available in digital format, and are often left out or underrepresented in global AI developments. This causes the resulting tools to be biased or less practical for local users, which in turn affects the frequency of usage.
Getting AI applications such as chatbots, learning apps, or digital services to communicate effectively in native languages is thus critical for mass adoption in many parts of the world where English is a second language for only the more educated population segments.
Looking ahead
Alibaba's Qwen 3 isn't just another AI model; it's a potent symbol of China's accelerating drive in a domain long led by American titans. Indeed, China has already superseded its Western rivals in some areas.
By delivering a competitively performing, open-source, and cost-effective tool that champions linguistic inclusivity across 119 languages, Qwen 3 is directly empowering developers in emerging markets, creating a more equitable global AI ecosystem.
For More Info on Asia Tech Lens