Alibaba Opens Qwen4-Max API to All Developers With 25% Price Cut Over Previous Flagship
The 2.4-trillion-parameter MoE model enters pay-as-you-go general availability on November 16.

Alibaba's Qwen4-Max is now generally available. As of November 16, 2026, developers can access the model through a standard pay-as-you-go API — ending the five-day Token Plan–only preview that began on November 11 — at pricing set 25% below the previous Qwen3.8-Max flagship.
The model is priced at $1.50 per million input tokens, $4.50 per million output tokens, and $0.075 per million cached tokens. That compares to $2.00 and $6.00 per million input and output tokens for Qwen3.8-Max, the model it effectively replaces at the top of Alibaba's commercial API stack. Access is available through Alibaba Cloud Model Studio and via OpenRouter under the model ID qwen/qwen4-max.
Inside the Architecture: A Hybrid Attention Design at Scale
Qwen4-Max is a sparse Mixture-of-Experts model with 2.4 trillion total parameters, activating 71 billion per token at inference. The design follows a 1-in-4 attention layering pattern: three out of every four layers use Gated DeltaNet (GDN), a linear attention mechanism that reduces compute by avoiding full quadratic attention; the fourth uses Qwen Sparse Attention (QSA), a standard sparse attention layer that maintains precision on complex reasoning tasks.
Each layer also incorporates a Gated Residual (GR) block with four parallel computation branches, adding representational capacity without increasing inference cost proportionally. The context window extends to one million tokens.
The most architecturally unusual component is N-gram Embedding, a 328-billion-parameter embedding table stored in CPU RAM and prefetched asynchronously during inference. Rather than relying solely on subword tokens, N-gram Embedding maps common token sequences — word pairs, frequent phrases, domain-specific compound terms — directly to embedding space. Alibaba's design intent is to improve retrieval on domain-specific text and reduce tokenization fragmentation effects for non-English languages.
This architecture was first described publicly by the Qwen team as the "Qwen3.8-Flash-Next" experimental preview when they introduced the hybrid attention approach ahead of Qwen4's full release.
Benchmark Results: Self-Reported Scores and the Independent Evaluation Gap
Alibaba's GA launch documentation reports a DeepSWE v1.1 score of 75.7 for Qwen4-Max, a result independently confirmed by The Next Web. DeepSWE measures autonomous software engineering — writing, running, and iterating on code to solve real GitHub issues without human checkpointing.
On Terminal-Bench 4.0, which tests autonomous command-line task completion including file navigation, shell scripting, and package management, Alibaba's GA blog reports a self-reported score of 48.5. That figure has not yet appeared on the Artificial Analysis independent leaderboard, where Qwen4-Max is not currently listed. The independent Terminal-Bench 4.0 leaderboard currently shows Kimi K3.1 Pro leading at 67.2%, with Claude Sonnet 5.5 at approximately 63.6%.
Alibaba also self-reports an Artificial Analysis Intelligence Index (AII) score of 43 for Qwen4-Max. The AII is a composite benchmark aggregating performance across reasoning, coding, math, and instruction-following tasks. Independent AII scores for competing top-tier models currently place Kimi K3.1 Pro at approximately 55 and Claude Opus 5.5 at approximately 58.
For the SWE-Marathon benchmark — which tests sustained, multi-session software engineering on complex repositories — no Qwen4-Max score has been published.
The gap between self-reported and independently verified benchmark results is not unusual at a model's GA launch. Independent evaluators typically require additional time to run full harness evaluations. Developers building production systems should treat Alibaba's GA-day figures as preliminary until third-party results are published.
Apple Intelligence in China: December Rollout Powered by Qwen
Qwen4-Max's GA arrives alongside a broader strategic development for Alibaba. Apple announced in October 2026 that Qwen will power Apple Intelligence features in China — including Siri responses, Writing Tools, and Visual Intelligence — beginning with the public release of iOS 27.3 in December 2026. A developer beta launched November 4.
The integration makes Alibaba the designated AI partner for Apple's China market, mirroring the arrangement with OpenAI for US and international markets. It represents a significant distribution channel for Qwen4-Max beyond direct API customers: Apple shipped more than 50 million iPhones in China in 2025, and iOS 27.3 will enable Qwen-powered features by default for users who opt in to Apple Intelligence.
Data Sovereignty and Security Considerations
Qwen4-Max is a product of Alibaba, a company headquartered in Hangzhou, China. Organizations in regulated industries or jurisdictions with data-residency requirements should evaluate their use accordingly.
China's National Intelligence Law requires domestic organizations to cooperate with state intelligence work when legally compelled. Decree 790, the Network Data Security Regulations that entered into force on January 1, 2026, establishes additional data-handling requirements for companies processing data related to Chinese national security interests.
Alibaba Cloud offers infrastructure in regions outside China, and enterprise customers can negotiate data-processing agreements and regional deployment options. Neither geographic deployment nor contractual terms fully eliminates the legal-access considerations created by a company's country of incorporation; organizations with strict compliance requirements should conduct their own legal review.
On AI safety, prior research published in a May 2025 paper evaluated closed-source large language models on multilingual jailbreak prompts and found that Qwen-Max — an earlier generation flagship — was the most susceptible to those attacks among the models studied. No independent red-team evaluation of Qwen4-Max's safety properties has been published as of this writing.
Open-Source Ecosystem and Developer Tooling
Alibaba has consistently released open-weight Qwen models alongside its commercial API offerings. The Qwen team maintains an active open-source ecosystem on GitHub, including developer tools and libraries that complement both commercial and self-hosted deployment of Qwen-family models. Qwen4-Max itself is a commercial API-only product; open-weight variants at different capability tiers are available separately through the Qwen repository.
What Comes Next
The move to full pay-as-you-go general availability positions Qwen4-Max for enterprise pipeline integration beyond the developer preview audience. With pricing set below comparable frontier models and Apple Intelligence distribution entering its final testing window, November 16 represents a meaningful commercial milestone for Alibaba's AI business.
Independent benchmark evaluations on Terminal-Bench 4.0, SWE-Marathon, and the Artificial Analysis Intelligence Index will determine whether Qwen4-Max's self-reported scores hold under third-party scrutiny — and where it ultimately lands in the increasingly competitive ranking of frontier coding and reasoning models.



