Alibaba Launches Qwen4-Max, a 2.4-Trillion-Parameter Flagship Pitched on Cheaper Agentic Work
Alibaba Launches Qwen4-Max, a 2.4-Trillion-Parameter Flagship Pitched on Cheaper Agentic Work
Alibaba unveiled Qwen4-Max on Sunday, November 15, its new flagship artificial intelligence model and the first release in the Qwen4 family. The company is pitching the model less on raw capability than on cost: it says the average price of completing a professional task falls by 62 percent compared with its predecessor, Qwen3.8-Max, while new training safeguards aim to keep the model away from offensive cybersecurity uses.
The launch follows a short preview that opened on November 11 on Alibaba’s Token Plan, with early access in QwenWork, Qoder, the Qwen app, and Qwen Studio. The full release adds API access and published pricing.
What Qwen4-Max is
Qwen4-Max is a sparse mixture-of-experts model built on a new Qwen4 architecture. It keeps the 2.4-trillion-parameter backbone of its predecessor, Qwen3.8-Max, but activates about 71 billion parameters per token instead of about 95 billion. The model supports a 1-million-token context window and has native vision, meaning it can work with images as well as text.
Alibaba positions the model for creative and professional work spanning 3D modeling, game development, precision manufacturing, finance, law, healthcare, and frontier scientific discovery, and says it outperforms Qwen3.8-Max, unveiled in August 2026, while using fewer tokens and costing less to run.
The cost claim, in numbers
The headline figure comes from Work360, Alibaba’s own benchmark, which evaluates performance on challenging industry tasks across 721 specialized fields. On that evaluation, Alibaba says Qwen4-Max uses 35 percent fewer tokens than its predecessor at High effort reasoning mode, and the average cost per task falls from $4.43 to $1.67, a reduction of about 62 percent.
The lower task cost reflects two changes. First, the model uses fewer tokens per task. Second, Alibaba cut its list prices: Qwen4-Max costs $1.50 per million input tokens and $4.50 per million output tokens, down from $2 and $6 for Qwen3.8-Max.
That pricing is aggressive. For comparison, Anthropic lists Claude Opus 5.5 at $20 per million output tokens, which underlines how much of this launch is a price play: Alibaba is betting that a cheaper model that is good enough for long, multi-step work will win developers who run agents all day.
In a manufacturing demonstration, the company said Qwen4-Max reduced manual workpiece programming time from approximately 475 minutes to 24 minutes, using a workflow that generates designs and checks machining plans through simulation.
Safety is part of the pitch
The launch puts unusual emphasis on safety, and not only in the usual sense of content filters. Alibaba says it filtered the model’s training data by task purpose for cybersecurity: it kept data for defensive tasks such as vulnerability verification, detection, and remediation, while removing data intended to improve offensive attack capabilities.
Code security checks added during training lifted the share of functions passing both functionality and security tests on the Last Secure Code Bench from 22 percent to 31.8 percent, the company says, and internal testing found no evidence of significant high-risk capabilities.
That emphasis fits a wider pattern. Chinese AI developers have increased safety disclosures in new model releases in recent months, amid growing global concern about so-called rogue AI and loss-of-control risks.
Reasons to stay cautious
Every performance and cost figure in this launch comes from Alibaba’s own evaluations, and independent testing has not yet had time to catch up. Sanchit Vir Gogia, chief analyst at Greyhound Research, said enterprises should treat the cost figure as a vendor benchmark: the relevant measure is the cost of an accepted result, including retries, reasoning tokens, and employee time spent checking outputs.
Yash Kadakia, founder and CTO of Security Brigade, made a similar point about the security claims, questioning whether offensive and defensive cybersecurity knowledge can be reliably separated through training-data filtering, since understanding an attack such as SQL injection is necessary both to exploit it and to prevent it. Both analysts recommended independent validation, for example from government evaluators or established external red teams, before treating the safeguards as settled.
Why it matters
Qwen4-Max arrives as the flagship-model race shifts from capability to economics. The largest labs now ship broadly competent models, so the differentiator is increasingly what a long agentic job costs and how reliably it finishes. A 62 percent cut in cost per task, if it holds up outside Alibaba’s own benchmark, would change which workloads are worth automating.
The safety framing matters too. By publishing training-data filtering and benchmark numbers for defensive capability, Alibaba is setting a precedent that other model makers may be pressed to match, and giving enterprise buyers a concrete set of claims to test. Whether those claims survive independent evaluation will decide if Qwen4-Max is a genuinely cheaper workhorse or simply a well-priced announcement. One limit is access: Qwen4-Max itself is not open source. It is available as a hosted service through QwenCloud, on the Token Plan or through the API, as well as in QwenWork, Qoder, and Qwen Studio. Alibaba says smaller models in the Qwen4 family are coming to open source, but the flagship weights stay closed, so outside researchers cannot verify the technical claims directly.
