Alibaba Just Launched Qwen4-Max at 2.4 Trillion Parameters, and the Real Story Is a Cost Per Task That Fell More Than 60 Percent
The flagship is big, the pitch is cheap
Alibaba unveiled its new flagship model, Qwen4-Max, on Sunday, November 15, in a launch reported by Reuters the same night. A week after its preview, the model went live on the QwenCloud API, in Qwen Studio, QwenWork and Qoder, and within hours it also appeared on OpenRouter and OpenCode Go.
The headline size is 2.4 trillion parameters. That number needs its usual translation. Qwen4-Max is a sparse mixture-of-experts model, which means it does not use all of those parameters on every token. Launch summaries put activation at about 71 billion parameters per token. The model carries a very large pool of specialized capacity, and pays the computing cost of only the slice each request actually routes through.
That design is the whole strategy in one sentence. Reuters framed the wider race plainly: Chinese technology companies, already a major force in open-weight models globally, are locked in a fast-moving battle to build more powerful systems without making them prohibitively expensive to run. Qwen4-Max is Alibaba’s latest answer to that exact problem, and most of its launch is about cost, not size.
What Qwen4-Max actually is
The model is multimodal with native vision. It handles text and images, supports a context window of up to one million tokens, and is built for long professional and scientific work rather than short chat.
Alibaba listed the work it is aiming at in its launch statement, quoted by Reuters: creative and professional tasks spanning 3D modeling, game development, precision manufacturing, finance, law, healthcare, and frontier scientific discovery. Launch demos followed that list, taking a 3D blockout to a cinematic cut, a sketch to industrial production, and designing new coolant materials.
Access follows the pattern buyers now expect from a serious challenger:
- API first. Qwen4-Max is available through QwenCloud, and Alibaba made it compatible with the OpenAI and Anthropic protocols. For a team already built on either interface, trying it is a configuration change, not a rewrite.
- First-party surfaces. The model is also live in Qwen Studio, QwenWork and Qoder, Alibaba’s own chat, work and coding products.
- Aggregators on day one. Listings on OpenRouter and OpenCode Go appeared within hours of launch, which puts the model in front of developers who never visit Alibaba’s own cloud.
- A preview window already running. A Token Plan preview opened on November 11 and runs until November 25, so early users were testing the model before the formal unveiling.
One limit matters for the open-weight audience. Qwen4-Max itself stays closed. Alibaba says smaller models in the Qwen4 series will follow as open weights, and reporting around the launch notes a Qwen4-27B small enough to run on a machine with 16GB of RAM. That split, a closed flagship with open smaller models below it, is now the standard Chinese playbook: earn revenue and control at the top, build the developer base underneath.
The efficiency numbers are the launch
Alibaba says Qwen4-Max outperforms its predecessor, Qwen3.8-Max, released in August, while using fewer tokens and costing less. The specific claims, all from Alibaba and launch coverage, stack up like this:
- About 50 percent fewer tokens. On Terminal-Bench 4.0, the model uses roughly 48 percent fewer tokens than its predecessor for comparable work, and launch reporting puts the overall token reduction at about half.
- Cost per task down sharply. On Work360, Alibaba’s own benchmark covering 26 domains, 182 subdomains and 721 specialized fields, Qwen4-Max scores 52.69, about 10 percent above Qwen3.8-Max, while average cost per task falls from $4.43 to $1.67. That is a drop of more than 60 percent.
- List prices cut. Pricing is $1.50 per million input tokens and $4.50 per million output tokens, with cached input at $0.075 per million, a caching price down about 70 percent from the previous generation.
These two savings multiply. A lower price per token is one cut. Using roughly half as many tokens to finish the same task is a second, independent cut on top of it. A buyer does not need Qwen4-Max to be the single smartest model in the world for that math to matter. It needs to be close enough on the work that matters, at a total task cost competitors cannot match.
How it scores, and whose numbers these are
On Alibaba’s own scorecard, Qwen4-Max posts 75.7 on DeepSWE v1.1, the real-repository coding benchmark, which the company says puts it ahead of Claude Opus 5.5 and competitive with Moonshot’s Kimi K3.1 Pro. It scores 48.5 on Terminal-Bench 4.0. Alibaba compared the model against Qwen3.8-Max, Opus 5.5, GPT-6.1 Sol, GPT-6 Astra and Kimi K3.1 across its launch charts.
Independent scorers are more measured, and their picture is still strong:
- Artificial Analysis gives Qwen4-Max 43 on its Intelligence Index, nine points above Qwen3.8-Max and three points behind Kimi K3.1 Pro, at about $3.27 per task against $3.68 for Kimi.
- Arena ranks it sixth in Code Arena for web development, with a score of 1750, up 74 points, and fourteenth on Agent Arena, the top Chinese model on that board.
- Vals puts it eleventh overall.
Read those together and the honest summary is narrower than the launch headline. Qwen4-Max is not, on independent boards, the top model in the world, and on Artificial Analysis it still trails Kimi K3.1 Pro. It is, on those same independent boards, the top or near-top Chinese model on several coding and agent rankings, with a cost per task below the rival it trails. That is a credible flagship claim, and it does not need the company scorecard to stand up.
A market footnote points the same way. Alibaba’s Hong Kong shares rose 1.2 percent in Monday morning trading, suggesting investors treated this as a real product step.
Safety was part of the release, not an afterthought
The most unusual part of the Reuters report is how much of it is about safety. Alibaba said Qwen4-Max’s training included security checks alongside coding tests, aimed at reducing vulnerabilities such as code injection, memory-safety flaws and broken access controls in the code the model writes. Safeguards against misuse in areas such as cybercrime and hazardous materials were strengthened, and the company said internal testing found no evidence of significant high-risk capabilities.
Two cautions belong with that. First, these are Alibaba’s own statements and its own internal testing. No independent safety evaluation was published with the launch. Second, the context Reuters gave matters: Chinese model developers have increased safety disclosures in recent releases amid growing global concern about rogue AI and loss-of-control risks. A safety section in a flagship launch is becoming a competitive requirement, not a voluntary extra, because enterprise buyers in finance, law and healthcare, the exact sectors Alibaba named, now ask for it before they ask about price.
Even with those cautions, the direction is worth noting. A model that writes production code is a security surface. Testing the code it generates for injection and access-control flaws during training treats that surface as part of the product, and buyers should expect every flagship launch from here to make the same claim and then be asked to prove it.
Why this launch matters beyond one model
Qwen4-Max lands in a crowded three weeks. Moonshot released Kimi K3.1 Pro in late October, billed at 2.8 trillion parameters. Meituan released the open weights for LongCat-2.5, a 1.6-trillion-parameter agent model trained entirely on Chinese chips, just yesterday. Alibaba’s entry is the largest of the three by total parameters, the only one that stays fully closed at the top, and the one whose launch leans hardest on cost per finished task rather than a single benchmark crown.
That last choice is the signal. The frontier race spent two years competing on the highest score. Recent launches compete on a different question: what does a finished, correct task cost, in tokens, in dollars, and in security cleanup afterwards. On that question, a model that scores 75.7 on DeepSWE in its maker’s tests, ranks near the top of independent coding boards, and cuts its predecessor’s task cost from $4.43 to $1.67 is a serious answer, even after company numbers are discounted.
The bottom line
Alibaba launched Qwen4-Max on November 15: a 2.4-trillion-parameter mixture-of-experts flagship, about 71 billion parameters active per token, a one-million-token context, native vision, OpenAI and Anthropic compatible APIs, and pricing at $1.50 per million input tokens and $4.50 per million output tokens. Alibaba reports 75.7 on DeepSWE v1.1 and a cost per task down more than 60 percent from Qwen3.8-Max. Independent trackers put it at 43 on the Artificial Analysis Intelligence Index, three points behind Kimi K3.1 Pro, at a lower cost per task. The flagship stays closed, smaller Qwen4 models are promised as open weights, and the safety claims in the launch are the company’s own, not yet independently verified. Test it on your own work before believing any scorecard, including this one. But on price, efficiency and day-one availability, Qwen4-Max is the flagship every Western lab now has to answer on cost.