Alibaba and Tencent bet on cheaper AI models for businesses
Alibaba launches Qwen4-Max, while Tencent unveils Hy4 flash, two models designed to reduce usage costs. The two Chinese groups target professional productivity, with announced gains in price, speed, and certain business applications.
Alibaba and Tencent each presented, on November 16, 2026, an AI model designed for professional use, with a clear priority: lowering the bill. Qwen4-Max at Alibaba and Hy4 flash at Tencent target long tasks, documents, spreadsheets, slides, and office automation, in a market where the price per task now matters as much as the quality of responses. Both announcements also show the same strategy: keeping a high-end model but offering a faster and cheaper version for daily use.
Deux modèles, deux promesses de productivité à moindre coût
At Alibaba, Qwen4-Max is presented as the new flagship model intended for professional and scientific applications. The group announces a 62% reduction in average cost per task, a price lowered to 1.50 dollars per million input tokens and 4.50 dollars per output million tokens, and a context window that can reach one million tokens. The model is meant to serve finance, health, law, industrial manufacturing, and scientific research.
Tencent follows a similar logic with Hy4 flash, a lighter version of Hy4 launched after the main model on November 4, 2026. The company says it has lowered prices by about 92% on input tokens and 89% on output tokens compared to Hy4, with rates of 0.070 dollar per million input tokens and 0.281 dollar per output million tokens. In short, the race is no longer just about "who answers best," but about "who answers well enough without burning a hole in the wallet."
Des architectures lourdes, mais davantage optimisées
Qwen4-Max relies on a Mixture-of-Experts architecture, or MoE, a system that activates only part of the model's parameters for each request. Alibaba mentions 2.4 trillion parameters in total, with 71 billion activated per token, which is 25% less than Qwen3.8-Max-Base. The model also includes native visual understanding, allowing it to process text and images in the same request.
On Tencent's side, Hy4 flash has 224 billion parameters in total, with 12 billion activated per token. The company says it has reduced the size of the key-value cache, which stores previous context computations for reuse, to one twenty-fourth of Hy3's size and one-sixth of Hy4's. It adds that all service optimizations, including megakernels, four-bit quantization, and DFly, enable decoding speeds exceeding 150 tokens per second.
Pourquoi les benchmarks internes comptent autant que les prix
Both groups seek to shift the debate towards measures closer to real work. Alibaba highlights Work360, its internal benchmark covering 721 specialized domains, and claims that Qwen4-Max outperforms Qwen3.8-Max. Tencent, for its part, built WorkIndex around concrete office tasks, such as slide creation, document editing, spreadsheet repair, audio/video processing, and visual understanding.
On WorkIndex, Tencent says Hy4 flash performs slightly better than Hy4 preview overall, while dividing the average cost per task by 13 compared to Hy4 preview, by 20 compared to Hy4, and by 8 compared to DeepSeek-V4.1-Flash-1017 at Max-tier peak hour pricing. Alibaba also provides an average cost calculation, from 4.43 dollars to 1.67 dollars per task, but these results have not yet been confirmed by independent evaluations at the time of the announcements [to be verified].
La sécurité et l’intégration deviennent des arguments commerciaux
Alibaba emphasizes enhanced security controls during Qwen4-Max training. The group says it has limited risks related to code injections, memory security flaws, and access control issues. It also claims to have excluded certain data associated with cyber-offensive activities while retaining those related to vulnerability detection and correction.
Tencent does not emphasize security as much in its communication but focuses on distribution. Hy4 flash is already available in Tencent Cloud TokenHub, WorkBuddy, Yuanbao, WeCom, QQ, ima, and Marvis, with a planned arrival on OpenRouter. Developer access uses the hy4-flash identifier, supporting OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages formats, as well as vision, reasoning, structured output, and function calling features.
Une bataille de gamme qui ressemble à un triathlon
The two launches outline a fairly clear hierarchy. At Alibaba, Qwen4-Max targets complex tasks and research, with a clear focus on heavy professional use. At Tencent, Hy4 flash rather targets volumes, short deadlines, and agents that chain actions without requiring a doctorate in patience. Both companies keep a more powerful model in parallel for difficult cases.
This segmentation says something about the current market: companies want power, but above all, they want to know how much it costs per line, document, or task. Generative AI is entering a less spectacular, more prosaic phase, where usage economy becomes a selling point as important as model size. And that is often where the real match is played.
Key points
- Alibaba launches Qwen4-Max on November 16, 2026.
- Tencent presents Hy4 flash on November 16, 2026.
- Qwen4-Max announces 62% average cost reduction per task.
- Hy4 flash shows up to 92% decrease on input.
- Work360 covers 721 specialized domains at Alibaba.
- WorkIndex tests real office tasks at Tencent.
By the numbers
- 2.4 trillion — total parameters of Qwen4-Max, according to Alibaba.
- 224 billion — total parameters of Hy4 flash, according to Tencent.
- 1 million tokens — context window announced by both models.
- 150 tokens per second — decoding speed claimed by Tencent.
- 85 users and 259 tasks — Tencent's internal test on WorkBuddy.