Alibaba Unveils Qwen4-Max, Its Largest AI Model Yet, With a 2.4-Trillion-Parameter Design and a 62% Cut in Task Costs
Alibaba has unveiled Qwen4-Max, its new flagship artificial intelligence model, pitching it as a tireless teammate for coding, office work and long-running agent tasks rather than as another chatbot. The launch, announced on Sunday and opened to developers on a pay-as-you-go basis on Monday, is the Chinese technology giant’s biggest model release yet, and it arrives with an unusually blunt commercial message: the model is designed to make each unit of AI work cost far less than it did on its predecessor.
According to Reuters, Alibaba said the new system improves its ability to handle complex professional and scientific work while strengthening protections against safety risks. The company says Qwen4-Max outperforms Qwen3.8-Max, the flagship it unveiled only in August this year, while using fewer tokens and costing less to run. Dow Jones reported that the model carries 2.4 trillion parameters and has achieved top rankings across leading benchmarks, particularly in coding, agentic tasks and visual reasoning.
The release comes at a moment when Chinese AI developers are moving fast and in the open. Beijing-based Moonshot AI released its Kimi K3.1 Pro model last month, calling it its most capable system to date with three trillion parameters, according to Dow Jones. Alibaba’s answer is not simply a bigger model. It is a cheaper one to operate, aimed squarely at the developers and companies now building software agents that call a model dozens or hundreds of times to complete a single job.
There is also a break with tradition in this launch, and it matters. Qwen4-Max itself is not open-weight. Developers cannot download its weights and run it themselves. Alibaba says smaller models in the Qwen4 family will be released openly later, but it has given no dates or sizes, and the flagship will remain a hosted service available only through its cloud platform. For a model family that built its global reputation on open releases, that is a significant change of direction.
What Alibaba actually launched
Qwen4-Max is the first model of Alibaba’s new Qwen4 generation and its new flagship. It went into preview on 11 November, available only to subscribers of Alibaba’s Token Plan and through its own apps, including QwenWork, the Qoder coding platform, the Qwen app and Qwen Studio. The general release that followed on 16 November adds metered, pay-per-token access through the QwenCloud application programming interface, with the model listed under the identifier qwen4-max.
The model is built on a new sparse mixture-of-experts architecture. According to Alibaba’s launch post, as reported by Dow Jones and specialist outlets, it has a backbone of 2.4 trillion parameters, of which only about 71 billion are activated for each token it processes, plus a further 328 billion parameters devoted to a new N-gram embedding design. That structure is the heart of the efficiency story: a very large model that only wakes a small fraction of itself for any single piece of work.
It is also a multimodal system. It accepts text, images and video as input, and it reasons by default, with thinking modes that can be set to low, high or xhigh effort. Its context window runs to one million tokens, meaning it can hold roughly a whole codebase, a long legal file or months of documents in view at once. The programming interface accepts the same formats used by OpenAI and Anthropic, so developers can switch models without rewriting their software, and it supports function calling, structured output, web search, a code interpreter and the Model Context Protocol for connecting tools.
Alibaba describes the model as a teammate for work, creativity and discovery. In its own words, quoted by Reuters, it “can take on a wide range of creative and professional work, spanning 3D modeling, game development, precision manufacturing, finance, law, healthcare, and frontier scientific discovery.” The company showed the model building three-dimensional scenes in the Blender design tool, making playable games in the Unity and Unreal engines, running financial research with traceable sources, and driving a 48-hour autonomous laboratory run that designed and synthesised a new coolant molecule.
The numbers behind the model
The headline figures all come from Alibaba itself, and they should be read that way. No independent laboratory had published a full evaluation of the model at the time of launch, and the early independent leaderboards that did publish results on launch day told a more mixed story than the company’s own charts. That gap is part of the story, not a footnote to it, and this article keeps the two kinds of evidence separate throughout.
On Alibaba’s own account, Qwen4-Max is a large step up from Qwen3.8-Max. The company says it outperforms its predecessor while using as much as 50 per cent fewer tokens on some tests, according to Reuters. On its internal Work360 evaluation, a benchmark Alibaba built to test tool use and long-running professional tasks across 26 domains, 182 subdomains and 721 specialised fields, Qwen4-Max scored 52.69, roughly 10 per cent higher than the 47.74 scored by Qwen3.8-Max. Token use fell by about 24 per cent at the highest reasoning effort and about 35 per cent at high effort, and by roughly 48 to 49 per cent on Terminal-Bench 4.0, a test of terminal-based coding work.
The combined effect of cheaper prices and fewer tokens is the figure Alibaba puts at the centre of its launch. On Work360, the company says, the average cost of completing a task fell from 4.43 dollars on Qwen3.8-Max to 1.67 dollars on Qwen4-Max, a reduction of about 62 per cent. That is a striking claim, and it is also a claim built on Alibaba’s own test, using Alibaba’s own prices. Independent analysts who examined the arithmetic noted that the 25 per cent price cut combined with a 35 per cent token saving accounts for roughly half the reduction, and that reaching 60 per cent or more requires token use to fall by about half, which Alibaba’s own figures do show on some tests but not on all of them.
On the company’s own benchmark chart, Qwen4-Max leads its rivals on some tests and trails them on others. On DeepSWE, a demanding software-engineering test, it scored 75.7, ahead of Anthropic’s Claude Opus 5.5 at 74.2, OpenAI’s GPT-6.1 Sol at 75.2 and Moonshot’s Kimi K3.1 at 75.5. On a legal benchmark, HLAB-Vals, it scored 21.3, far ahead of Claude Opus 5.5 at 9.2. But on Terminal-Bench 4.0 it scored 48.5, behind Claude Opus 5.5 at 59.6 and Kimi K3.1 at 67.2. On ProgramBench it scored 31.5, against 65.0 for Claude Opus 5.5. On Humanity’s Last Exam, a hard general-knowledge test, it scored 51.7, behind Claude Opus 5.5 at 59.8. On APEX-Agents it scored 72.1, just behind Claude Opus 5.5 at 73.5. The pattern is consistent: strong on legal and long-horizon agent work, competitive on software engineering, and behind the leaders on some coding and reasoning tests.
The first independent results, published on launch day, sharpened that picture. Artificial Analysis, which runs an independent intelligence index, scored Qwen4-Max at 43 points at its highest reasoning effort, nine points ahead of Qwen3.8-Max and three points behind Kimi K3.1 Pro at 46. On the Code Arena WebDev leaderboard, the model debuted in sixth place with a preliminary 1,750 points, 74 points above Qwen3.8-Max. On Agent Arena it debuted fourteenth overall, up from thirty-ninth for its predecessor, and it was the top-ranked Chinese model on that board. Those scores are early and will move as more votes and test runs come in, but they already show a model that is genuinely near the front of the field without leading it.
What it costs to use
The price list is one of the few parts of the launch that is a matter of public record rather than a company claim. On QwenCloud, Qwen4-Max costs 1.50 dollars per million input tokens and 4.50 dollars per million output tokens, with cached input priced at 0.075 dollars per million tokens. Against Qwen3.8-Max, which costs 2 dollars for input and 6 dollars for output, that is a 25 per cent cut on both, and the cached-input price has fallen 70 per cent, from 0.25 dollars.
Those prices matter because of how the model is meant to be used. A chatbot answers a question once. An agent plans a task, calls tools, checks the result, retries when something fails, and calls the model again, sometimes dozens of times before the job is done. Each of those calls costs tokens, and the bill for a long task can grow quickly. By cutting both the price per token and the number of tokens each task needs, Alibaba is attacking the total cost of a completed job, not just the price of a single answer.
Artificial Analysis put that in perspective on its own index. It calculated that an average task on its index costs about 3.27 dollars on Qwen4-Max, against about 10.04 dollars on Claude Opus 5.5. Only OpenAI’s GPT-6.1 Sol was cheaper, at about 1.27 dollars. That comparison is not a full cost-of-ownership calculation, and it says nothing about quality, but it explains why Alibaba chose cost per task as the headline of its launch. For a company running agents at scale, a model that is three times cheaper per task is a serious proposition even if it is not the smartest model available.
There are also demonstrations aimed at buyers rather than benchmark readers. Alibaba says the model cut the manual programming effort for a single workpiece at two automotive suppliers from about 475 minutes to 24 minutes. In a legal demonstration, it sorted 198 working papers for pre-IPO due diligence and delivered an 86-page report. Both are company demonstrations, not outside validation, and they should be treated as illustrations of what the model is being sold for, not as proof of what it will do in a buyer’s own environment.
A closed flagship from the open-weight champion
The most consequential decision in this launch may be the one about who gets to own the model. Qwen4-Max is close-sourced. Its weights are not available for download, and there is no release on Hugging Face or GitHub. Alibaba says smaller models in the Qwen4 family will be released as open-weight later, allowing developers to download and modify them, but it has given no sizes, dates or licence terms, and the flagship itself will stay closed.
That is a break with the strategy that made Qwen one of the most widely used model families in the world. Earlier Qwen generations were released openly, and developers, researchers and small companies built on them freely. The open releases were also a source of soft power for Alibaba and for China’s AI sector as a whole, demonstrating capability in a form the whole world could inspect and use. A closed flagship keeps that capability behind Alibaba’s own paywall.
The commercial logic is straightforward. A hosted model earns money every time it is called. An open-weight model earns goodwill and adoption, but the revenue goes to whoever hosts it, which is often someone else. As the cost of training frontier models climbs into the hundreds of millions of dollars, the pressure to charge directly for the best model grows. OpenAI, Anthropic and Google have all kept their flagship models closed from the start. Alibaba is now joining them at the top of its own range, while keeping the open-weight tradition alive further down the range, where smaller models can still be downloaded and run locally.
Whether that trade-off pays off will depend on what the smaller open models look like when they arrive. If they are strong, Alibaba can keep both its developer community and its paying customers. If they are weak, or slow to arrive, the company risks losing the community that made Qwen a household name among developers. The Qwen team has said the smaller models are coming to open source, but the absence of a timetable leaves the question open.
Safety moves to the front of the announcement
Alibaba gave safety an unusually prominent place in this launch. According to Reuters, the company said the model’s training included security checks alongside coding tests, helping reduce vulnerabilities such as code injection, memory-safety flaws and broken access controls. Safeguards to prevent misuse in areas such as cybercrime and hazardous materials were strengthened, and internal testing found no evidence of significant high-risk capabilities, the company added.
That emphasis is part of a wider pattern. Reuters noted that Chinese AI developers have in recent months increased the safety disclosures in new model releases, amid growing global concern about rogue AI and loss-of-control risks. The timing matters. Qwen4-Max is being sold for coding agents, which is exactly the use case where misuse risk is highest: a model that can write and run code, call tools and operate for hours without supervision is also a model that could be turned to harmful ends if its safeguards fail.
Alibaba published one concrete safety figure. On the Last Secure Code Bench, which tests whether generated code is both working and secure, the share of code passing both tests rose from 22.0 per cent on Qwen3.8-Max to 31.8 per cent on Qwen4-Max. That is a real improvement, and it is also a reminder of how far there is to go: on that measure, more than two-thirds of the code the new model writes still fails one test or the other. Safety, on Alibaba’s own evidence, is improving from a low base.
The company’s safety claims are self-reported, and no outside audit has been mentioned. That is normal for a product launch, but it means the claims should be treated as a statement of intent rather than a verified result. Independent testing of the model’s refusal behaviour, its resistance to prompt injection and its handling of dual-use requests will be one of the more useful things outside researchers can do in the weeks after launch.
The road from preview to launch
Qwen4-Max did not appear out of nowhere. Alibaba said at its Apsara conference in Hangzhou on 22 September that its next-generation model, Qwen 4, was in training. The Qwen team announced on 11 November that Qwen4 was coming and that Qwen4-Max was in preview on its Token Plan subscription, with early access in QwenWork, Qoder, the Qwen app and Qwen Studio. The preview ran for two weeks and was due to end on 25 November.
The preview itself was thin on detail. The announcement promised deeper knowledge, richer world understanding and stronger coding and agentic capabilities, but it contained no benchmark scores, no per-token price and no parameter count. Specialist outlets that examined Alibaba’s documentation at the time found that the model was not yet listed in the public model pages, and that the newest entry in the Qwen Cloud changelog was from 27 October. By the standard those outlets applied, a model is not really released until a script can call it and its price can be read, and on that standard Qwen4-Max was not released until the general availability announcement on 16 November.
That staged rollout is now a familiar pattern in the AI industry. A preview builds anticipation and gathers early feedback from friendly users, while the full launch supplies the numbers that let buyers and rivals judge the product. The five days between preview and general availability also gave Alibaba time to publish its benchmark chart and price list, which are the two documents the market actually needed.
Why now — the race Qwen4-Max enters
The launch lands in the middle of a fast-moving contest between Chinese and American AI developers, and between Chinese developers themselves. Moonshot’s Kimi K3.1 Pro arrived last month with three trillion parameters, and Alibaba’s own Qwen3.8-Max arrived only in August. The pace of releases has become so fast that a flagship model can be superseded within a quarter, and the competition is no longer only about which model is smartest. It is about which model is cheapest to run at scale, which one developers can build on most easily, and which one can be trusted with long, unsupervised tasks.
The contest also plays out against a background of export controls and geopolitical tension. The United States has restricted sales of advanced AI chips to China since 2022, and the controls have been tightened repeatedly since then. NEXA has reported extensively on the fallout, including the resignation of Nvidia’s chief executive this month and a Reuters review that identified more than three billion dollars in Chinese computing projects built around Nvidia’s restricted processors. Chinese developers have responded by building more efficient models that do more with less computing power, and by developing domestic chip capacity. Qwen4-Max, with its emphasis on doing more work with fewer tokens, fits that strategy precisely.
There is a parallel story in the United States. OpenAI launched a new science blog this week, built around work in which its GPT-6 Astra model helped a physicist develop a candidate mathematical definition of the Standard Model of particle physics. The work was posted as a preprint in October and featured as the blog’s first post on 15 November. It is a different kind of announcement from Alibaba’s, aimed at researchers rather than buyers, but it points in the same direction: the leading AI companies are now competing as much on scientific discovery and long-horizon work as on chatbots.
Alibaba’s own position in this race is distinctive. It is one of China’s largest cloud providers, which means it can sell Qwen4-Max directly to the businesses that already run on its infrastructure. It is also, through the Qwen family, one of the most important suppliers of open AI models in the world. Qwen4-Max tests whether those two roles can still be held together: a closed flagship that earns money, and an open family that earns trust.
Analysis: Why It Matters
The price is the product
Alibaba could have led this launch with intelligence. It chose to lead with cost. That choice tells us where the AI market actually is in late 2026. The frontier models from the leading American and Chinese laboratories are now close enough in capability that, for most commercial work, the deciding question is not which model is smartest but which one delivers a completed task at the lowest price. A model that finishes a job for 1.67 dollars instead of 4.43 dollars changes the economics of every business built on top of it, and it changes them in a way that a few extra benchmark points do not. Alibaba understands that, and it has aimed Qwen4-Max directly at the bill rather than the leaderboard. The risk in that strategy is that a price-led launch invites a price war, and Alibaba’s rivals, including Moonshot and the American laboratories, have shown no reluctance to cut prices when their own costs fall. But for now, cost per task is the most honest measure of what an AI model is worth to a buyer, and Qwen4-Max is the first flagship launch to put it at the centre of the announcement.
A closed flagship from the open-weight champion
The decision to keep Qwen4-Max closed is the most important strategic signal in this launch, and it deserves to be read carefully. Qwen’s open releases did more than win developers. They gave Alibaba a form of influence that money cannot buy: researchers around the world studied its models, built on them, and taught the next generation of engineers with them. Closing the flagship trades some of that influence for revenue, and it does so at a moment when open-weight models from Chinese rivals are becoming more capable, not less. The promise of smaller open Qwen4 models softens the break, but a promise is not a release. If those models arrive quickly and perform well, Alibaba will have squared the circle. If they arrive late or weak, the company will have given up its most distinctive advantage in exchange for a hosted service that competes on price with every other hosted service. The next few months will show which of those outcomes is real.
The benchmark split is the honest picture
Read Alibaba’s benchmark chart without the marketing gloss and it tells a coherent story. Qwen4-Max is strong where long, structured, professional work matters: legal analysis, long-horizon agent tasks, software engineering at scale. It is weaker where raw reasoning and terminal-based coding are measured, trailing Claude Opus 5.5 and Kimi K3.1 on Terminal-Bench and ProgramBench by wide margins. The independent leaderboards agree: a top-six coding debut, a top-fifteen agent debut, and an intelligence index score behind Kimi K3.1 Pro. That is not a failure. It is a model with a clear profile, built for a particular kind of work, and the honest reading is that buyers who need long-running agents and professional document work will find it compelling, while buyers who need the hardest coding and reasoning tasks will still look elsewhere. The companies that survive this race will be the ones whose customers understand that distinction, and Alibaba has, to its credit, published enough of the unfavourable numbers for a careful reader to draw it.
Safety as a selling point, not a footnote
The prominence of safety in this launch is new, and it is welcome. For years, safety disclosures in model releases were brief, vague and buried at the end of the announcement. Alibaba has put security testing, misuse safeguards and a concrete secure-coding figure near the front of its launch, and Reuters noted that other Chinese developers have been doing the same. There are two ways to read that. The generous reading is that the industry has learned that trust is a feature, and that buyers of agent systems need to know what a model will refuse to do before they let it run unsupervised. The sceptical reading is that safety disclosure has become a competitive signal, a way to look responsible while the underlying capabilities keep growing. Both readings can be true at once. What matters now is whether independent researchers are given the access to test these claims properly. A safety claim that cannot be checked is a slogan. A safety claim that survives independent testing is an asset. Qwen4-Max’s safety story will be decided by the second of those tests, not the first.
The race is now fought on efficiency, not just scale
Step back from the individual numbers and the shape of the contest becomes clear. The first phase of the AI race was about scale: who could build the biggest model. The second phase, which Qwen4-Max embodies, is about efficiency: who can do the most work with the fewest tokens, the least computing power and the lowest bill. That shift suits Chinese developers, who have been forced by export controls to squeeze more from less, and it suits every buyer who has watched AI bills climb. It also changes what counts as progress. A model that is ten per cent smarter but twice as expensive is no longer obviously better than a model that is ten per cent cheaper and nearly as smart. Qwen4-Max is a bet that the market has already made that judgement, and that the next frontier is not a smarter chatbot but a cheaper, steadier worker that can run for hours without supervision. If that bet is right, the launch will be remembered as a turning point. If it is wrong, it will be remembered as a competent model that arrived a little early for its own argument.
What to Watch Next
- Independent benchmark updates. Artificial Analysis, Code Arena and Agent Arena will revise their Qwen4-Max scores as more test runs and votes come in. Watch whether the model climbs toward the leaders or settles where it debuted.
- The smaller open-weight Qwen4 models. Alibaba has promised open releases of smaller models in the family, with no dates or sizes. Their timing and quality will decide whether the closed flagship costs Alibaba its developer community.
- Enterprise adoption. The car-manufacturing and legal demonstrations are Alibaba’s own. Watch for named customers and case studies from companies outside China, which would show the cost-per-task argument working in the real world.
- Rival responses. Moonshot, OpenAI, Anthropic and Google all price against each other. A price cut or a new efficiency-focused release from any of them would show how seriously they take the Qwen4-Max challenge.
- Independent safety testing. Outside researchers will test the model’s refusal behaviour and its resistance to prompt injection and misuse. Their findings matter more than the company’s own safety figures.
- Alibaba Cloud’s results. The next quarterly results will show whether Qwen4-Max is converting developers into paying cloud customers, and whether the model is driving growth in Alibaba’s cloud division.
- The export-control question. Qwen4-Max was built under American chip restrictions. Any change in those controls, or any new evidence about how Chinese developers are obtaining computing power, would change the context in which this model was built.
Sources
- Reuters — Alibaba unveils new flagship AI model with safety upgrades
- Dow Jones via TradingView — Alibaba Unveils AI Model With Stronger Performance at Lower Cost
- KWSN — Alibaba unveils new flagship AI model with safety upgrades (Reuters)
- AI Europe — Alibaba’s Giant Qwen4-Max Model Trails Claude, GPT and Moonshot AI
- The Model Atlas — Qwen4-Max is out: price, benchmarks and how to try it
- Crypto Briefing — Alibaba unveils Qwen4-Max, claims it beats rival models on coding