Alibaba Releases Qwen3 as Open-Weight Model with State-of-the-Art APAC Multilingual Performance

Alibaba Cloud has released Qwen3 — the third generation of the Qianwen large language model family — as open-weight models under the Apache 2.0 licence, with the Qwen3-72B flagship model achieving state-of-the-art performance on APAC multilingual benchmarks including C-Eval (Mandarin reasoning), CMMLU (Mandarin knowledge), Japanese NLP benchmarks, and Korean language understanding evaluations, while performing comparably to GPT-4o on English reasoning benchmarks.

Qwen3's APAC language performance — particularly on Mandarin Chinese reasoning and knowledge tasks — reflects Alibaba's advantage in APAC language model training: Alibaba's access to Chinese web corpora, Chinese e-commerce and business document data, and Chinese academic content through its commercial operations provides training data scale and quality for Chinese-language model capability that US model developers cannot easily replicate. Qwen3-72B's C-Eval score of 94.2% significantly exceeds GPT-4o's 87.2% on the same benchmark — a performance gap that matters for APAC enterprises deploying Chinese-language AI applications.

Qwen3's release under Apache 2.0 — the most permissive open-source licence, allowing commercial use, modification, and deployment without royalty or usage restrictions — enables APAC enterprises to self-host Qwen3 on their own APAC cloud infrastructure for Chinese-language AI applications without per-token API cost or US-hosted data dependency. For APAC financial services companies processing Chinese-language documents, APAC e-commerce companies generating Chinese-language product content, and APAC enterprises building Chinese-language customer service AI, Qwen3 provides a self-hostable alternative at GPT-4o-comparable or superior Chinese-language quality.

The broader APAC open-source AI implication of Qwen3's release: the combination of Llama 4 (Meta, strong English/multilingual), Qwen3 (Alibaba, strong APAC/Chinese), and Mistral Large (European regulatory alignment) creates an open-source model portfolio that APAC enterprises can evaluate against commercial APIs for self-hosted deployment, substantially improving APAC enterprise AI sovereignty options.