Alibaba said to revamp AI app to rival ChatGPT

Alibaba is planning to revamp its main mobile AI app, aiming to make it more similar to OpenAI’s ChatGPT, according to sources cited by Bloomberg.

The Chinese ecommerce giant will update its existing “Tongyi” apps on iOS and Android, renaming them “Qwen” after its AI model, and gradually add features to support shopping on platforms such as Taobao.

The company reportedly has over 100 developers working on the project, with plans to eventually launch an international version.

Alibaba operates several consumer AI apps, including Qwen Chat, and intends to unify the user experience under the Qwen brand.

The Qwen app will remain free for now, but Alibaba may seek to monetize consumer-facing services in the future.

China’s leading tech firms, including Huawei and Tencent, are increasing investments in AI to compete with US companies like OpenAI and Meta.

Alibaba has also updated its Quark search app with AI features, which will continue to be available.

🔗 Source: Bloomberg

🧠 Food for thought

Implications, context, and why it matters.

The LMSYS Chatbot Arena (a community leaderboard by the Large Model Systems Organization with human A/B voting) uses a Bradley–Terry model (pairwise ranking) mapped to an Elo-like scale (chess rating). In that setup, Qwen2 placed around position 15 while Meta’s Llama-3 70B (a 70-billion-parameter model) ranked near 11 1 2.
AAII v3 (a composite benchmark suite for large language models) includes MMLU-Pro (a tougher version of Massive Multitask Language Understanding), GPQA Diamond (graduate-level science questions), and coding challenges 3. Developers and venture investors can cross-check LMSYS, CanAiCode, and Big Code Models 4. This gap matters for production uses like support or shopping assistants, since it affects retention and API switching costs.
In Alibaba Cloud Model Studio (Singapore) which is a managed service for deploying and testing models, Qwen-Plus has a non-thinking mode for up to 256K input tokens per request. Input runs about $0.4 per million tokens and output about $1.2 per million tokens (tokens are billable chunks of text the model reads or writes) 5. Groq lists Qwen3-32B at roughly $0.29 per million input tokens 6.
Multi-turn chats raise token use because history counts as input each turn. Developers can cut costs by about 50% with Batch Inference (processing requests asynchronously in bulk, Singapore region only). Cache hits get billed at 10% of the standard input token price, and savings plans from $10 to $5,000 add predictability 7. The free quota is available only in the Singapore region 7.