EMPVResearch NotesIT
EMPV / RESEARCH NOTESNOTE / 024

NEWS NOTE / AI MODELS

Qwen4-Max: Alibaba targets more efficient AI agents with 71 billion active parameters

The new model combines 2.4 trillion total parameters, a one-million-token context window and capabilities for complex workflows. Alibaba reports a 58% reduction in average task cost compared with Qwen3.8-Max, based on its proprietary benchmarks. The model remains cloud-hosted.

Published2026-11-19 ยท 5 min
Qwen4-MaxAlibabaAI agentsMixture of ExpertsAI APIsAI safety
01 / Architecture

A 2.4-trillion-parameter architecture

On November 16, 2026, Alibaba introduced Qwen4-Max, the new flagship model in the Qwen family.

The release focuses on reasoning efficiency and the use of external tools. Qwen4-Max is designed for multi-step tasks where an agent must retrieve information, interact with software and produce a usable result.

Alibaba reports improvements across coding, professional work and tool-use evaluations. Its launch demonstrations include industrial and document-intensive workflows. Those results still need to be distinguished from independently reproducible performance in production environments.

Qwen4-Max uses a Mixture-of-Experts (MoE) architecture with 2.4 trillion total parameters, activating approximately 71 billion parameters per token.

The distinction matters. Total parameters describe the model's overall capacity, while active parameters indicate how much of the network is involved in each generation step.

Compared with Qwen3.8-Max, Alibaba reports 25% fewer active parameters while retaining the same overall model scale.

The updated architecture combines Gated DeltaNet, Qwen Sparse Attention, gated residual connections and N-gram embeddings. These mechanisms are intended to improve processing efficiency, particularly for long contexts.

Qwen4-Max supports a context window of up to one million tokens, accepting text, images and video as input and generating text as output.

A large context window allows more information to be provided in a single request. It does not, by itself, guarantee accurate retrieval or use of every detail within that context.

02 / Benchmark

Work360: stronger results with fewer tokens

According to Alibaba Cloud's Work360 evaluation, Qwen4-Max scores 52.69, compared with 47.74 for Qwen3.8-Max.

Work360 is a proprietary benchmark covering 26 primary domains, 182 subdomains and 721 specialized categories of professional work.

Alibaba also reports 35% fewer tokens used at the High reasoning-effort setting and a 58% reduction in average cost per task.

That reduction should not be interpreted as a universal 58% discount on API calls. Task cost depends on token pricing and the amount of computation required to complete the work.

At launch, Alibaba also reported a sixth-place ranking on Code Arena's web development leaderboard.

These figures help identify capabilities worth evaluating, but they do not establish universal agent reliability. Performance depends on the task, the available tools and the evaluation conditions.

03 / Industry

From demonstrations to industrial workflows

Alibaba accompanied the release with demonstrations across manufacturing, scientific research, legal work and financial analysis.

In one manufacturing example, Qwen4-Max used engineering information to build three-dimensional production-line representations, generate CAM toolpaths and identify machining issues through simulation.

In the case shown, manual programming time per part fell from approximately 475 minutes to 24 minutes, a reduction of about 95%.

Another demonstration connected the model to 90 tools through the Model Context Protocol (MCP). The agent collected and cross-checked financial information, producing an Excel workbook and an editable presentation with references to its sources.

These examples illustrate a broader engineering direction: connecting language models to specialized tools and coordinating work across multiple stages.

They remain selected demonstrations. They do not establish that similar savings can be consistently reproduced across factories or administrative departments.

04 / APIs and pricing

APIs, pricing and availability

Qwen4-Max is available through QwenCloud and Alibaba Cloud Model Studio, using the `qwen4-max` API identifier.

It supports function calling, parallel tool calls, web search, structured outputs and interfaces compatible with OpenAI and Anthropic API protocols.

According to Model Studio documentation updated on November 18, 2026, standard pricing in the Frankfurt region is listed at $1.413 per million input tokens and $4.24 per million output tokens.

Terms vary by region, caching and processing mode. The actual cost of an agent also includes repeated model calls, external tools, infrastructure and any required human verification.

**Qwen4-Max is not distributed with publicly downloadable model weights.** Alibaba has announced plans to release smaller open-weight Qwen4 models, without a confirmed date.

The separately released Qwen-Live-Delivery-Plugin is open-source tooling for iterative agent workflows. It supports intermediate deliveries and user feedback during long-running tasks, but does not make Qwen4-Max an open-weight model.

For companies processing sensitive information, the availability of a Frankfurt endpoint does not by itself establish that all data processing remains within the European Union. Applicable data residency, retention and contractual conditions still need to be reviewed.

05 / Security

Security and enterprise integration

Alibaba also reports improvements in secure code generation. On Last Secure Code Bench, the proportion of outputs meeting security checks simultaneously at function, file and repository level increased from 8% for Qwen3.8-Max to 14.7% for Qwen4-Max.

Reuters notes that the company's new safety claims have not been independently verified by the agency.

For businesses, model-level security and end-to-end system security are separate concerns.

Consider an agent that receives a customer order through a message and must enter it into an existing ERP. It needs to identify the customer, validate products and quantities, respect commercial conditions, handle failures and prevent unauthorized operations.

Understanding the message is only one part of the workflow. Reliable execution also requires access controls, data validation, operational logs and recovery procedures.

Qwen4-Max expands the options available to engineers building such systems. Whether it is suitable for a particular business process still needs to be established through representative tests measuring errors, total cost, human intervention and the reliability of the final outcome.

EMPV / TAKEAWAY

Qwen4-Max offers new options for AI agents integrated with business systems. Vendor benchmark results should be compared with errors, costs and controls measured on real workflows.

EMPV / SHARE

Share this Research Note.

Research Notes