阿里千问发布 Qwen4-Max:2.4T 参数、百万 token 上下文,主打长时程智能体
Qwen4-Max: Your New Teammate Has Arrived
阿里千问发布 Qwen4-Max,基于新的 Qwen4 架构,拥有 2.4 万亿参数主干、原生视觉能力和一百万 token 上下文窗口。官方称其在 Work360 基准上得分从 Qwen3.8-Max 的 47.74 提升到 52.69,Work360 单任务平均成本从 4.43 美元降至 1.67 美元,降幅 62%。模型已在 QwenCloud 上线,可通过 Token Plan 或 API 使用,同时接入 QwenWork、Qoder 和 QwenStudio,官方表示 Qwen4 系列更小的开源模型即将推出。
推荐理由
AI 评分 88
官方披露了 2.4T 参数新架构、百万 token 上下文与 token 成本下降幅度,可据此判断长任务智能体的落地条件。
主题
正文
QWEN STUDIO DISCORD
Today, we are introducing Qwen4-Max, your new AI teammate for work, creativity, and discovery. Qwen4-Max is built on the new Qwen4 architecture, with a 2.4-trillion-parameter backbone, native vision capabilities, and a one-million-token context window.
Qwen4-Max advances real-world productivity through greater proficiency with professional tools, stronger long-horizon task execution, and higher-quality deliverables. 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. As a 24/7 AI teammate, Qwen4-Max keeps the big picture in view and moves projects forward, bringing ambitious ideas closer to reality and opening new ways for people and AI to work together.
Comprehensive performance comparison across coding, work, and long-horizon tasks. Qwen4-Max represents a significant step up from Qwen3.8-Max, delivering stronger performance across nearly every capability we evaluated. See Footnotes for evaluation details.
Token efficiency and cost. Qwen4-Max outperforms Qwen3.8-Max with fewer tokens: on Work360, token usage drops by approximately 24% and 35% at Xhigh and High effort, respectively; on Terminal-Bench 4.0, the reductions are approximately 48% and 49%, respectively. We have also lowered API prices compared with Qwen3.8-Max, with particularly large reductions in cached-input pricing. In our Work360 evaluation, the average cost per task drops from $4.43 to $1.67 — a 62% reduction.
Pre-training#
Qwen4-Max is built on a new architecture for the Qwen4 series, with 2.4T backbone parameters, 71B activated per token, and an additional 328B N-gram Embedding parameters. The architecture systematically upgrades the sparse MoE model across attention, residual connections, and embeddings, improving computational efficiency, model capacity, and training stability:
GDN + Qwen Sparse Attention (QSA): Three Gated DeltaNet layers compress history into fixed-size states for every QSA layer. QSA uses a lightweight indexer to select relevant context at micro-block granularity, reducing both attention and indexing costs for long sequences.
Gated Residual (GR): Four parallel residual branches use dynamic read and write gates to strengthen information flow across layers and improve training stability.
N-gram Embedding: Context-based lookups provide representations of local token patterns at the second layer, expanding model capacity with little additional computation.
Qwen4-Max architecture.
Training at the 2.4T-parameter scale progressed smoothly, with loss decreasing steadily and no loss spikes throughout the run. We attribute this stability primarily to improvements in model architecture and training optimization.
Smooth and stable pre-training of Qwen4-Max.
On in-house evaluations across Knowledge, Code, QA, and STEM, Qwen4-Max-Base achieves the lowest bits per byte (BPB) among all five evaluated checkpoints. It improves on Qwen3.8-Max-Base in every domain, with approximately 5–13% lower BPB while activating 25% fewer parameters.
In-house evaluations: 10 × BPB, lower is better. GLM-5.3 and Kimi-K3.1-Pro are post-trained checkpoints, which may affect comparisons with base models.
Real-World Productivity#
Qwen4-Max brings stronger creative and professional capabilities to complex workflows across industries, delivering high-quality, usable results—your new AI teammate for real-world productivity.
Industrial and Professional Work#
To strengthen the model’s collaborative capabilities across industrial and professional workflows, we developed Work360, a comprehensive framework for modeling and categorizing complex work scenarios. Its hierarchical taxonomy spans 26 Tier-1 domains, 182 Tier-2 subdomains, and 721 Tier-3 specialized fields, providing broad coverage with fine-grained distinctions across industries and professions. Building on this taxonomy, we developed the Work360 benchmark to evaluate performance on challenging industry tasks. In the latest evaluation, Qwen4-Max scored 52.69, up from 47.74 for Qwen3.8-Max, marking a substantial improvement over the previous generation. Qwen4-Max combines broader industry coverage with stronger professional task execution and collaboration, making it a more capable and versatile tool for real-world productivity.
3D Modeling and Generation#
Qwen4-Max demonstrates strong 3D creation capabilities. It uses Blender to carry out geometry modeling, scene assembly, material setup, lighting, and rendering, progressively turning creative ideas into detailed 3D works with a distinct visual style. Across architecture, interiors, natural landscapes, and science-fiction scenes, it demonstrates an integrated understanding of spatial layout, object structure, and visual appearance.
These 3D assets are in practical used in games and brought into physical objects through 3D printing. This illustrates the model’s potential to connect digital creation with downstream production workflows, supporting design exploration, content creation, and prototyping.
Qwen4-Max works fluently with leading game engines, including Unity, Unreal Engine, and Godot. From a single natural-language brief, it handles scene construction, gameplay logic, integration of suitable external assets, material setup, HUD design, and visual effects to deliver a playable game. During development, it runs the game and uses visual and interaction feedback to refine camera tracking, character animation, hit feedback, lighting, and pacing. Through repeated testing and iteration, it brings scenes, gameplay, and sound together into a cohesive 3D game experience.
Meow of Magic
Qwen4-Max built a cat-mage adventure in Unreal Engine. Players cast spells, choose upgrades, and take on waves of enemies and bosses.
Yun Guan Xing
Qwen4-Max built a fantasy adventure in Unreal Engine. Players fly on a sword through clouds and rings, explore the world, and face a boss.
Flying Racer
Qwen4-Max built a mobile racer set around West Lake in Unity 6. Players drift, use nitro boosts, and deploy wings to fly in an installable Android game.
In real production environments at SANF and Pacific Union, Qwen4-Max demonstrates end-to-end capabilities spanning automotive production-line planning, part design, and machining verification. By connecting design generation, tool execution, and feedback-based validation, it delivers measurable efficiency gains.
At the production-line level, Qwen4-Max uses 2D layouts and process parameters to build 3D production-line models from an existing equipment library, simulate cycle times, and perform PLC virtual commissioning. At the part level, it models and designs car-door assemblies, friction wheels, and other components, while assisting with machining process planning, toolpath generation, and simulation-based verification. Through feedback-driven iteration, it identifies overcuts and insufficient material removal before machining begins. In this showcase, the manual effort required to program a single workpiece fell from approximately 475 minutes to 24 minutes, a reduction of about 95%.
Agentic Video Production#
With Qwen4 and WonderClip, you can build an end-to-end video creation workflow — from refining storyboard scripts and constructing static assets to animating cameras in blockouts, rendering them into individual clips, and editing them into a cohesive, multi-scene film.
The powerful language, multimodal, and agentic capabilities of Qwen4 bring knowledge integration, controllability, and stability to complex video creation. Knowledge is incorporated into generative models through text prompts or reference images, helping the generated content better reflect the creator’s knowledge and reasoning. Parametric blockouts ensure consistent geometry, scale, and spatial relationships across complex static assets. They also enable precise control over camera paths and the timing of actions and events. Natural-language interaction, combined with controllable previsualization, makes the final results more predictable and interpretable.
Qwen4-Max supports complex work in high-barrier scenarios such as financial investment research, legal due diligence, and clinical healthcare. Beyond processing sprawling raw data, it executes deep logical deduction and evidence-chain traceability—truly achieving a leap from “information retrieval” to “professional-grade delivery.”
Financial investment research: Qwen4 adeptly invokes 90 MCP tools to comprehensively acquire market quotes, financials, interest rates, and dividend data across A-shares, HK and US equities, bonds, funds, and derivatives, while integrating search to cross-check public disclosures. For more complex research tasks, it tracks industry supply-demand dynamics and changes in new orders, assesses macro interest rates and credit spreads, evaluates the equity-bond yield spread, and conducts multi-factor backtesting to gauge style exposures.
In the CSI 300 index derivatives case, it combines financing costs, dividend yields, and contract maturities to calculate the theoretical futures basis, analyze the actual discount to spot, and examine the spread between implied and historical volatility. Ultimately, deliverables are generated as Excel workbooks and editable PowerPoint decks, backed by dual traceability mechanisms: first, verifiable sources, where clicking hyperlinks within cells leads directly to the original data source; and second, chart-to-workbook alignment, where PowerPoint charts map one-to-one to the underlying Excel workbook, enabling analysts to trace back from the presentation to verify exact figures and formulas.
Pre-IPO legal due diligence: Qwen4-Max demonstrates exceptional ability to handle complex, long-horizon legal work. Faced with 198 inconsistently formatted and named working papers supplied by the client in batches, the model effectively filters out noise, renames and categorizes the files by substantive content into 14 folders, and automatically compiles a working paper index. It also proactively searches public records covering business registration, intellectual property, judicial records and credit records, cross-checking these sources against the working papers to verify facts concerning the issuer. For core risk control, the model scrutinizes contractual provisions and analyzes legal risks, accurately identifying potential concerns such as reinstatement clauses in valuation adjustment mechanism (VAM) agreements and providing actionable remedial recommendations. Ultimately, it delivers a comprehensive report meeting rigorous professional standards: 86 pages, over 60,000 Chinese characters, 12 review areas including corporate history and shareholding structure, and 14 specific legal issues. Every conclusion is backed by working paper citations for full traceability, drastically boosting efficiency and empowering legal counsel to focus on core risk judgments.
Critical care: Patients often present with multiple complications and rapidly changing clinical status. Frequent diagnostic testing generates a massive volume of heterogeneous medical data—dozens of lab reports, over 10,000 Chinese characters of progress notes, multiple sets of physician orders, and 1,433 original medical images. The patient in this video developed ventricular fibrillation after carotid artery stenting and was successfully resuscitated, only to experience chest pain post-extubation and require reintubation on Day 6. Within a single conversation, Qwen4 reviews the entire complex medical record, completing an end-to-end workflow from information extraction to clinical decision support: querying lab values (accurately locating specific CRP results across multiple dates), structuring the clinical course (aligning six days of complex events into a coherent timeline), interpreting imaging (categorizing studies into six imaging series, pinpointing the target chest CT, and deploying multiple subagents to simultaneously review hundreds of original slices to identify bilateral dependent consolidation, fully consistent with the attending physician’s diagnosis), and supporting decisions (synthesizing multimodal clinical evidence to propose four management recommendations that correspond point by point with the physician’s actual orders that day). Through close integration into the clinical workflow, the model efficiently consolidates evidence from multiple sources to provide reliable, actionable support for frontline clinical decision-making.
Long-Running Agents (24/7 Agents)#
Qwen4-Max powers 24/7 agents that tackle long-running tasks through autonomous planning and continuous iteration, delivering usable results in stages while keeping users in control.
When users have to wait until a long-running task is finished to see the results, an agent’s mistakes can be difficult to undo. Qwen4-Max’s Live Delivery keeps long-running tasks on track through continuous delivery and real-time user guidance. The agent delivers a usable first version early, then expands and improves it through successive releases, updating its plans based on execution results and user requirements. Each release gives users a concrete basis for assessing progress and deciding what should happen next. They can step in at any point to correct mistakes or change direction, staying in control as the work progresses.
Qwen4-Max is capable of operations research modeling, solving, and validation around complex objectives and constraints, thereby delivering high-quality optimization decision solutions. This capability has been deployed in real-world Alibaba Cloud business scenarios, achieving a closed loop in which the model autonomously iterates on and optimizes the very scheduling solutions it requires, specifically encompassing multi-model call scheduling for data annotation, online allocation and scheduling of cloud computing GPU resources, and offline rescheduling of cloud computing CPU resources.
In this process, the model autonomously generates candidate algorithms based on task specifications, collaborates with MindOpt, Alibaba’s proprietary solver, to evaluate feasibility, objective values, and time consumption to obtain feedback, and then iterates toward superior strategies by modifying code and performing cross-validation. From upper-layer model invocation to underlying cloud resource scheduling, Qwen4-Max continuously evolves based on quantifiable feedback, directly optimizing the operational systems that support its own applications, inference, and training. The resulting cost reductions and improvements in resource efficiency, in turn, provide more robust computing power to support the next round of model iteration. In these respective scenarios, model call scheduling costs for data annotation were reduced by an average of 10.42% compared with the reference baseline, online GPU allocation increased whole-machine idle capacity duration by 4.62%, and in the offline CPU rescheduling scenario, the solving speed for reaching the optimal objective improved by more than 14.61 times.
Promoting Scientific Discovery#
To address the cooling challenges that next-generation high-power chips pose for data centers, Qwen4 used the integrated dry- and wet-lab environment of Deep Principle’s Mira platform to autonomously take a new coolant molecule from design through laboratory synthesis to testing and validation, ultimately yielding a physical sample of the material.
During a continuous 48-hour agent run, Qwen4 worked entirely on its own: it surveyed possible technical approaches, constructed a candidate space of 2.7 million molecules, and then proceeded through structural screening, MPA property prediction, first-principles calculations, retrosynthetic analysis, and experiment planning. Qwen4 went on to direct an automated wet lab to prepare reagents and carry out synthesis, separation and purification, and testing, turning molecular structures on a screen into physical samples in the lab. After four iterations of a closed “molecular design → synthesis and testing → active learning” loop, Qwen4 obtained a new, experimentally validated coolant molecule.
From AI for materials to materials for AI infrastructure, this work explores a new materials-based path toward tackling the data-center cooling problem and demonstrates Qwen4’s capability for long-horizon scientific discovery that bridges the digital and physical worlds.
Alzheimer’s disease progressively impairs memory and cognition, placing a long-term care burden on families and society. Neuroinflammation is an important component of its pathology, and the NLRP3 inflammasome, which participates in this process, represents an important avenue for research into new intervention strategies. Focusing on this target, Qwen4-Max, working on DP Technology’s PharmMaster platform, demonstrates end-to-end capabilities spanning target research, molecular optimization and design, and computational evaluation. It goes beyond generating molecular structures, integrating literature and patent interpretation, specialized tool use, and computational feedback into an iterative scientific workflow.
In this process, Qwen4-Max autonomously searches literature and patent evidence, selects a reference molecule, and establishes an evaluation baseline. It then conducts multiple rounds of structural design, property prediction, and molecular docking, using evaluation feedback to refine its optimization strategy. Finally, it combines cross-round metric comparisons with relative binding free-energy (RBFE) assessment to prioritize a candidate. Across the workflow, it makes 454 tool calls, completes 3,239 computational metric evaluations, and designs and optimizes 24 candidate molecules—connecting evidence retrieval, design generation, tool execution, and feedback-driven verification into one end-to-end process.
Compared with the patent-derived starting molecule, the final prioritized molecule achieves broad improvements across multiple key computational metrics. RBFE calculations estimate its nominal binding affinity for the target at approximately 3.6 times that of the reference. This outcome has also received expert endorsement from DP Technology, whose experts consider the molecule to have substantial research value and to warrant further wet-lab validation. From a disease-associated target to a high-potential candidate molecule, Qwen4-Max combines evidence-driven design, specialized tool execution, and computational feedback, demonstrating its ability to autonomously advance long-horizon scientific research tasks, iteratively optimize molecules, and deliver valuable research outcomes.
Model-Harness-Product CoDesign#
Qwen4-Max is developed through close co-design across the model, agent harness, and product layers, in collaboration with the teams behind products such as Qoder. This approach defines model capabilities from real user workflows and evaluates them on end-to-end delivery, with a particular focus on everyday office work and professional design. By improving both task performance and the user experience, Qwen4-Max turns ideas into finished deliverables.
Safety#
At Alibaba, we embed model governance—guided by our six principles for AI development (human-centeredness, inclusivity and integrity, privacy protection, safety and reliability, trustworthiness and controllability, and open co-governance)—throughout the entire model lifecycle—from initial design and research through development, training, deployment, and real-world application—ensuring responsible AI is built in by design, not added on as an afterthought. For software engineering training tasks, we have added code security checks on top of functionality checks to strengthen the model’s intrinsic secure-coding capabilities. On Last Secure Code Bench’s “code reconstruction without security hint” and “dual verification of functionality and vulnerabilities,” Qwen4-Max’s function-level pass rate for both functionality and security rose from 22.0% for the previous generation (Qwen3.8-Max) to 31.8%; the share of code generated securely and simultaneously at all three granularities—function, file, and repository—rose from 8% to 14.7%, nearly doubling. The model performs particularly well at avoiding core risks such as injection, memory safety, and broken access control, markedly reducing vulnerabilities in code generated by default and helping developers shrink the system’s attack surface at the source. For cybersecurity, we filter data according to the specific purpose of each task, retaining data for defensive tasks such as vulnerability verification, detection, and remediation, while removing data for offensive tasks aimed at enhancing attack capabilities. A review by our internal security experts of long-horizon task trajectories (including full chain-of-thought, CoT) found that Qwen4-Max’s cybersecurity defense capabilities have improved substantially, that safety against offensive uses has also been significantly strengthened. No capabilities posing large-scale cybersecurity risks have been observed. For CBRN, we designed dedicated safety data and enhanced Qwen4-Max’s safety during the safety post-training stage. In internal evaluations, Qwen4-Max identified more requests carrying potential risks and provided safe responses, without exhibiting over-refusal. A review by internal experts of the model’s answers (including full CoT) found that the solutions the model proposes are not genuinely innovative and cannot complete the task end-to-end. No risk has been identified of the model translating relevant knowledge into a significant uplift in a user’s capabilities.
Ethics Statement#
Qwen is intended for lawful, constructive purposes and must not be used to cause harm. It cannot be used to exploit or endanger people, including children, or to promote violence, hate, discrimination, or abuse, or to develop or use CBRN weapons; nor may it be used to infringe privacy, carry out fraud or deception, spread disinformation, engage in cybersecurity abuse, or bypass security safeguards. Qwen is designed to assist people in their work, not to replace professional judgment in high-impact scenarios. Where decisions may carry serious consequences, human expertise and oversight remain essential. Users of Qwen must ensure that all input data is obtained and used lawfully, and that authorization has been obtained from the individuals concerned for any images or videos used. Users must not infringe individuals’ rights in their own images or videos or other personality rights, and must comply with applicable laws and regulations.
How to use#
Qwen4-Max is now available on QwenCloud. You can start building with either the Token Plan or the API.
You can also use the model on QwenWork, Qoder, and QwenStudio.
Summary#
Qwen4-Max is your 24/7 AI teammate for work, creativity, and discovery. Advanced reasoning, native vision, and long-horizon execution come together to turn complex goals into usable deliverables and tangible research outcomes. With Live Delivery and model–harness–product co-design, it keeps projects moving, adapts to your feedback, and hands back results you can review, trust, and build on. And Max is only the first look at Qwen4. Smaller models in the family are coming to open source, built on the same lineage and trained with the same generation of techniques, tuned for what small models do best. Your new teammate has arrived. Ready when you are.
Footnotes#
JobBench: Our scores were obtained using the updated dataset and evaluation code.
APEX-Agents v1.1: Scores are cited from the Mercor leaderboard.
AA-Briefcase v1.1 and AutomationBench-AA scores are cited from the Artificial Analysis.
CodeArena 和 AgentArena scores are cited from the LM Arena.
HLAB-Vals-1.1 scores are cited from the Vals Index.
Agents’ Last Exam v1.1 scores are cited from the Agents’ Last Exam.
ChartoGraphy: The scores for Qwen4-Max and Qwen3.8-Max-0902 were obtained with min_pixels set to 4096×32×32.
Citation#
@misc{qwen4, title = {Qwen4-Max: Your New Teammate Has Arrived}, url = {https://qwen.ai/blog?id=qwen4-max}, author = {{Qwen Team}}, month = {Nov}, year = {2026} }