Alibaba Cloud released a new version of its Qwen3.8 Max artificial-intelligence model on September 2, sharpening its coding, visual-analysis and tool-use capabilities as Chinese developers compete for business customers.
The release is an updated snapshot of the model Alibaba first unveiled on August 3, not the launch of an entirely new model family. Its identifier, qwen3.8-max-0902, records the date of the update.
According to Alibaba Cloud’s model documentation, the new version is intended to handle longer software-development projects, coordinate multiple tools and work with visual material such as charts and documents.
The update focuses on longer, multi-step work
Qwen3.8 Max has a context window of one million tokens. A token is a small unit of text, while the context window determines how much material a model can consider during one request. A larger window can accommodate lengthy documents or substantial parts of a software repository, although capacity alone does not guarantee accurate answers.
Alibaba also says the model can operate in thinking and non-thinking modes. Thinking mode allows the system to spend more computing time on a difficult request. The faster mode may be more economical for routine work.
Pricing has become part of the competition. On Alibaba Cloud’s standard Singapore price list, Qwen3.8 Max costs $2 per million input tokens and $6 per million output tokens. Both rates are 20% below the prices listed for Qwen3.7 Max in the same region. The comparison excludes promotional offers, caching discounts and batch processing.
Google provided another price reference on the same day as Alibaba’s update. It introduced Gemini 3.8 Flash at an introductory standard API price of $0.75 per million input tokens and $3.75 per million output tokens through December 31, 2026. Output billing includes tokens the model uses for reasoning.
Those rates are below the listed Qwen3.8 Max prices, but the products are positioned differently. Google describes Flash as a fast model for software engineering, AI agents and multi-step reasoning, while Qwen3.8 Max is Alibaba’s flagship model. Price alone cannot show which one completes a particular job more economically.
Input tokens cover information sent to the model. Output tokens cover the response it produces. The full cost of a business application also depends on how many attempts a task requires, how much supporting software is needed and whether employees must review the result.
Tencent and DeepSeek are pursuing similar work
Alibaba’s update follows a series of releases from other Chinese AI companies.
Tencent released and open-sourced Hy4 preview on August 28. The company says the model contains 770 billion parameters, with 49 billion activated for each token, and can process a context exceeding one million tokens.
This design is known as a mixture of experts. Instead of using every part of the model for every request, it activates a smaller selection, which can reduce the computing needed to generate an answer.
Tencent is offering Hy4 through its WorkBuddy and CodeBuddy products, its Yuanbao and ima applications, Tencent Cloud TokenHub and OpenRouter. Its stated uses include programming, office work, financial analysis and scientific research. Reuters independently reported the release and its main technical specifications.
DeepSeek brought DeepSeek V4 Pro into general availability on August 13. It also emphasized AI agents, adjustable reasoning effort and compatibility with OpenAI’s Responses API. That compatibility can reduce the work involved when developers test DeepSeek inside an application built around a familiar interface.
DeepSeek introduced separate peak and off-peak prices for the V4 family, with off-peak rates set 50% below peak prices. Scheduling flexible work for quieter periods gives businesses another way to control inference costs, the expense incurred when a trained model processes requests.
Business adoption will depend on the complete service
Alibaba, Tencent, DeepSeek and Google are now packaging their models around tasks buyers can recognize: reviewing long documents, writing and checking code, analyzing data and operating external tools.
Benchmark results remain difficult to compare. Model providers choose many of their own tests, software settings can affect the scores, and performance on a controlled evaluation may not carry across to a company’s documents or systems. Alibaba’s and Tencent’s performance statements should therefore be treated as company claims.
Businesses will also weigh reliability, response speed, data controls, regional availability and technical support. A cheaper model can become expensive if it requires repeated attempts or extensive human checking.
Open and adaptable models can still create opportunities beyond their original developers. As we reported in August, Thomson Reuters built its proprietary legal AI system from an existing open-source foundation that its chief technology officer later identified as Alibaba’s Qwen3.5.
That example shows how the commercial advantage can move beyond training the underlying model. A company can adapt an existing foundation with specialist information, product design and industry expertise that rivals cannot easily reproduce.
Alibaba’s September update gives developers another model to evaluate, but product documentation cannot determine a winner. Businesses will still need to measure accuracy, speed and total cost on their own work.
