Qwen3 Brings 119 Languages to Open-Weight AI — Alibaba's Bid for the Multilingual Race

The Facts

Alibaba's Qwen3 model family, released in April 2025 under the Apache 2.0 licence, has become one of the most widely used open-weight alternatives to proprietary frontier models, with benchmark performance that rivals or exceeds GPT-4-level capability across a range of standard evaluation tasks. The model family's particular strength in multilingual tasks — including Southeast Asian languages where most major AI models have historically underperformed — makes it exceptionally relevant for ASEAN developers building products for local language markets.

Disclosure: RECATOOLS is written with the assistance of Claude, made by Anthropic, which is also named in this article. Readers should weigh that interest.

Chinese AI labs have closed the capability gap with US leaders far faster than most Western analysts predicted. Alibaba's Qwen, ByteDance's models, and DeepSeek's reasoning-focused releases collectively represent a tier of capability that matches or exceeds many proprietary alternatives on standard benchmarks, with permissive licensing that enables commercial deployment.

The cost trajectory of capable AI model inference has compressed dramatically. Open-source models run on consumer and cloud hardware now deliver GPT-4-equivalent capability at inference costs below $0.10 per million tokens — a 10-100x reduction from the costs at which similar capability was first commercially available in 2023.

Technical Deep-Dive

Qwen3's multilingual capability reflects a deliberate expansion of its training data beyond English. Most large language models were trained predominantly on English-language data, with other languages underrepresented proportional to internet content ratios. ASEAN languages — particularly lower-resource languages like Malay, Indonesian, Thai, and Vietnamese — received minimal explicit investment in early model generations.

Qwen3 was trained on about 36 trillion tokens covering 119 languages and dialects, including Indonesian, Malay, Thai and Vietnamese. Alibaba has not published a breakdown of its Southeast Asian sources, so quality on regional idioms and on code-switching between English and local languages (a common feature of ASEAN digital communication) is worth testing before deployment.

Qwen3 includes two Mixture-of-Experts models alongside six dense ones. The flagship Qwen3-235B-A22B activates 22 billion of its 235 billion parameters per token; the smaller Qwen3-30B-A3B activates 3 billion of its 30 billion. Activating only a fraction of the parameters per token keeps inference cost closer to that of a much smaller model — the efficiency gain that enables competitive pricing.

The ASEAN Perspective

For ASEAN developers, Qwen 3's combination of multilingual strength and open-source permissive licensing creates a compelling option for applications serving local language users. A startup building a customer service chatbot for Indonesian users, a healthcare information app in Thai, or a financial planning tool in Bahasa Malaysia can deploy Qwen locally — maintaining full data residency, paying no per-token API fees, and serving users in their native language with genuine linguistic quality.