Technology•

Alibaba's Qwen4-Max Bets on Cheaper Coding Work With Half the Tokens and Lower Prices

Alibaba released Qwen4-Max on November 16, claiming a 75.7 DeepSWE coding score, roughly half the token use of its predecessor and task costs cut by more than 60%.

Written By :MINRK
Alibaba's Qwen4-Max Bets on Cheaper Coding Work With Half the Tokens and Lower Prices

Photo by Luca Bravo on Unsplash

1. Alibaba launches Qwen4-Max

Alibaba released Qwen4-Max on November 16, 2026, as the flagship of its Qwen4 family of artificial intelligence models. The company says Qwen4-Max scores 75.7 on the DeepSWE v1.1 coding benchmark, ahead of rival flagship models, while using about half the tokens of its predecessor and cutting average task costs by more than 60%.

The release continues a rapid cadence of model launches from the Chinese technology group, which has been competing with American and Chinese developers to offer capable large language models to businesses and software developers. With Qwen4-Max, however, Alibaba is placing less weight on claims of raw capability and more on cost, efficiency and practical usefulness.

That emphasis shows up across the launch: in the pricing, in the token-efficiency figures and in the protocol compatibility that is meant to make switching easy for developers already building on other platforms. Independent rankings, meanwhile, paint a more mixed picture of where the model stands on performance.

2. What is Qwen4-Max?

Qwen4-Max is Alibaba's top-tier large language model, built on a sparse mixture-of-experts design with a 2.4-trillion-parameter backbone. It accepts text, images and video as input and can work with a context window of up to 1 million tokens.

A mixture-of-experts architecture divides a model into many specialized sub-networks, often described as experts. Instead of activating the entire model for every request, a routing mechanism sends each piece of work to the subset of experts best suited to it. The approach allows a very large model to run more economically, because only a fraction of its parameters are in use at any given moment.

The context window describes how much information a model can consider at once, functioning as a kind of working memory. A window of 1 million tokens is large enough to hold lengthy contracts, research reports or substantial portions of a software codebase in a single session. For coding work in particular, that capacity can help the model understand how different parts of a large project fit together rather than looking at isolated files.

Support for images and video extends the range of tasks the model can handle. Developers can, for example, pass it screenshots, diagrams or recorded material alongside written instructions, which is useful in workflows that mix documents, visuals and code.

3. Access through QwenCloud and familiar protocols

Developers reach Qwen4-Max through an application programming interface on QwenCloud, Alibaba's platform for its Qwen models. The interface is compatible with the protocols established by OpenAI and Anthropic, two of the most widely used AI providers among developers.

That compatibility is a practical selling point. Many companies have already built their software around one of those providers' interfaces. Because Qwen4-Max can speak the same protocols, a team testing it may only need to change configuration settings, such as an endpoint address and a model name, rather than rewriting the code that calls the model.

Lowering that switching cost matters in a market where developers frequently compare models on price and quality. If trying an alternative takes minutes rather than weeks of engineering work, more teams are likely to run their own tests, which gives a cheaper model a better chance of winning business on cost.

Alibaba also ran a preview of the model through its Token Plan offering from November 11 to November 25, 2026, giving developers an early opportunity to try Qwen4-Max before and around its formal launch.

4. The benchmark claims and the fine print

Alibaba's headline performance figure is a score of 75.7 on DeepSWE v1.1, a benchmark designed to test how well models handle software engineering tasks. The company presents that result as evidence that Qwen4-Max outperforms competing flagship models on coding.

Independent rankings are less flattering. Third-party evaluations place Qwen4-Max behind some rivals, including models from Kimi, on overall performance. Those same assessments, however, give the model credit for its favorable cost per task, which is the area where Alibaba is concentrating its pitch.

The gap between vendor results and independent testing is common in the AI industry. Companies typically choose benchmarks and test settings that show their models at their best, while outside evaluators apply a common methodology across many models. Coding performance in particular can vary widely depending on the language, the size of the project and the kind of task involved, so a strong score on one benchmark does not guarantee the same advantage in every real-world setting.

5. How much cheaper is Qwen4-Max?

Qwen4-Max is priced at $1.50 per million input tokens and $4.50 per million output tokens, which Alibaba positions as cost-effective within the flagship tier. On top of the lower list price, the company says the model uses about 50% fewer tokens than its predecessor, Qwen3.8-Max, and that average task costs have fallen by more than 60% compared with that model.

Tokens are the small chunks of text that language models read and produce, roughly corresponding to pieces of words. Most AI services bill by the token, charging separately for the input a customer sends and the output the model generates. Output tokens usually cost more, because generating text requires more computation than reading it.

The two efficiency gains reinforce each other. A model that needs fewer tokens to complete a task reduces the bill even at a constant price per token. When the price per token also falls, the savings multiply. That compounding effect is why Alibaba can claim a reduction in task costs that is larger than either the cut in token usage or any change in pricing on its own.

For businesses running AI at scale, those numbers can be significant. A company processing large volumes of documents or running automated coding assistants across many developers may make millions of model calls, and a reduction of more than 60% in the cost of each task can change the economics of a project considerably.

The split between input and output pricing also matters for how different customers experience those savings. Workloads that feed the model long documents or large codebases and ask for short answers are dominated by input costs, where the $1.50 rate applies. Tasks that ask the model to write lengthy code or detailed reports lean more heavily on output tokens at $4.50 per million. Because Qwen4-Max is designed to reach answers with fewer tokens, the reduction in output volume can be especially valuable for the second group, since that is where each token is most expensive.

Lower costs can also widen the range of tasks that make sense to automate. Work that was previously too expensive to hand to a flagship model, such as reviewing every change in a large software project or processing an entire archive of records, can become viable when the price of each task falls sharply.

6. A shift from raw power to practical value

The Qwen4-Max launch reflects a broader change in how Alibaba is marketing its models. Rather than competing purely on who has the most capable system, the company is focusing on efficiency and practical applications, particularly professional productivity and coding.

Alibaba is targeting industries such as healthcare, finance and law. These sectors share two characteristics that suit its strategy: their work is heavy on documents, and cost control matters. Reviewing contracts, summarizing medical records or analyzing financial filings all involve processing large amounts of text, which makes the long context window and lower per-task costs directly relevant.

In such settings, a model does not necessarily need to top every leaderboard to be valuable. If it performs reliably on the specific tasks a business cares about and does so at a lower cost than alternatives, it can be the more attractive option even if a rival scores slightly higher on general benchmarks.

7. Open weights for smaller models, a closed flagship

Alibaba plans to release smaller models in the Qwen4 series with open weights, meaning developers will be able to download them, run them on their own hardware and adapt them for their own purposes. The flagship Qwen4-Max, by contrast, will remain closed and available only through Alibaba's services.

This two-track approach serves different goals. Open-weight models help build a wide community of developers who become familiar with the Qwen family, contribute improvements and build products on top of it. That broad adoption can, in turn, feed demand for Alibaba's paid services.

Keeping the largest and most capable model closed protects what is likely the company's most valuable AI asset. It also ensures that customers who want the best performance in the series must use Alibaba's own platform, where the company controls pricing and can generate revenue from each request.

8. What Qwen4-Max means for developers and businesses

For companies evaluating AI providers, the most important figures in the launch are the lower token usage and the lower per-token pricing. Together they make Qwen4-Max a candidate for workloads where cost has been a barrier, particularly high-volume tasks that run continuously.

Developers already using OpenAI or Anthropic interfaces face a low barrier to testing the model, thanks to the protocol compatibility. That makes it straightforward to run side-by-side comparisons on a company's own tasks, which is generally the most reliable way to judge whether a model is a good fit.

Buyers would be wise to weigh independent results alongside Alibaba's own benchmark claims, especially for coding, where performance can differ considerably from task to task. On the evidence available at launch, the efficiency case for Qwen4-Max looks stronger than the claim that it leads the field on capability. For many businesses, though, a model that is good enough and markedly cheaper may be exactly what they are looking for.

Related Articles

NEWSLETTERS

Don't miss another story.

Subscribe to the MINRK Newsletter today.

Tick the newsletters you want. No selection? You'll get MINRK Headlines.

By signing up, you will receive emails about MINRK products and you agree to our terms of use and privacy policy.

MINRK Daybook

Market analysis for crypto traders and investors.

EVERY WEEKDAY

MINRK for Advisors

Defining crypto, digital assets and the future of finance for financial advisors.

EVERY THURSDAY

MINRK Protocol

Exploring the tech behind crypto one block at a time.

WEEKLY

MINRK Long & Short

A must read for institutions. Insights, news and analysis delivered weekly.

EVERY WEDNESDAY

MINRK Headlines

The biggest crypto news and ideas of the day.

EVERY WEEKDAY

MINRK Policy

Examining the intersection of cryptocurrency and government.

WEEKLY

MINRK Research

Join thousands of readers who rely on MINRK for data-driven insights on the latest digital asset trends.

MONTHLY