Source type: Official release/benchmark results Publication: Official Qwen Blog Author: Qwen Team Publication date: 2026… This is a necessary excerpt; read the original source for full context.
Qwen positions Qwen3.8-Max as a flagship model for coding, office work, research, long-horizon tasks, and multimodal agents, claiming 2.4T total parameters, 95B activated parameters, reasoning effort, API access, and integration with multiple agent harnesses. The official article presents many internal benchmarks and end-to-end cases, with strengths concentrated in long-chain tool use, visual agents, document work, and continuous feedback; these figures are vendor-reported results and should be cited together with the evaluation framework, number of runs, and benchmark type.
Terminal Bench 2.1: 86.6; SWE-bench Pro: 67.7; PaperBench: 93.0; OSWorld-Verified: 86.1.
CoWorkBench: 74.8; WorkSpaceBench: 67.7; WideSearch: 81.9.
GPQA Diamond: 92.6; IFBench: 82.8; MRCR v2 256K (8-needle): 92.9.
MMMU-Pro: 82.3; MathVision: 95.2 / 97.7; VideoMME (with subtitles): 90.4.
Official cases: about 16 days of autonomous operation leaving 265 commits, 127 PRs, and 151 issues; about 125 continuous hours reproducing a paper, 7,600 lines of code, and 33 rounds of GPU training; WWW2025 competition accuracy rising from 0.60 to 0.853.
E-Commerce Bench: final cash of ¥416,252, which the official source says is 38% higher than second-place GLM 5.2 and 152% higher than Qwen3.7-Max.
The official article mixes internal benchmarks, public benchmarks, and demonstrative cases; different models also use different harnesses (such as Claude Code, Codex, OpenCode, or Qwen-Agent). The numbers in all tables therefore cannot be treated as a cross-model ranking under one set of experimental conditions. The article itself notes that some external results come from leaderboards or official reports.
Qwen Studio Qwen Code Research API Platform Ambassadors Simplified Chinese Download Use Qwen Studio Qwen3.8-Max: Coding and Office Work, a Comprehensive Leap 2026/08/03 · 43 min · 8572 words · QwenTeam | Translation: English
QWEN STUDIO DISCORD
Today, we officially release Qwen 3.8-Max, the most powerful model in the Qwen family to date. This also marks the first time that a Qwen-Max model will open-source its weights—the weights will be released next week, so stay tuned.
Qwen 3.8-Max is built on the architectural foundation of Qwen 3.5, with its parameter scale expanded to 2.4 trillion. It delivers comprehensive improvements in coding, office work, research, and long-horizon tasks. It can tackle more challenging problems, complete complex tasks end to end more reliably, and produce trustworthy results.
Qwen3.8-Max — now available on the Qwen AI Platform: Total parameters: 2.4T (95B activated); open-source weights will be released next week Comprehensively upgraded for coding, office work, research, and long-horizon tasks Reliable end-to-end delivery of complex tasks Callable directly through the API on the Qwen AI Platform. Coding capabilities
For frontier models, coding is no longer just “write a function as requested.” It now means starting from an empty folder and independently delivering a real project that spans several days. We tested Qwen3.8-Max on three such challenges—each result had to be produced by writing and running its own code, with no human assistance. One thread runs through all three cases: Qwen3.8-Max does not mechanically execute a fixed plan, but continuously evolves through feedback loops—building a self-upgrading harness, refining research methods through repeated experiments, and climbing a competition leaderboard through successive submissions.
Ten-day-scale autonomous coding: Building a self-evolving harness
In this case, Qwen3.8-Max was asked to create the oh-my-cli project from scratch and build a self-evolving harness through roughly ten days of long-horizon autonomous coding. It brought user feedback, advanced community practices, and model self-test results into one engineering loop, allowing requirements to be normalized into issues, automatically claimed and executed by an agent, and continuously iterated through code, tests, previews, and log feedback. The project's complete trace is publicly preserved in the GitHub repository qwen-code-dev-bot/oh-my-cli.
Key implementation details in building the autonomous coding harness:
Loop Engineering Setup: task state, dispatch, and recovery. Qwen3.8-Max combined an issue state machine, task dispatcher, monitor, and watchdog into one execution loop: after a new requirement entered a GitHub Issue, the agent automatically claimed it according to the state machine and moved through ready → leased → active; once implementation was complete, it triggered end-to-end tests and CI checks, then merged the PR after they passed. Self-testing: product self-testing and maintenance. After submitting an update, the model triggered Build, Unit Test, E2E, and Desktop Lifecycle validation; abnormal states automatically returned to issue/PR repair and re-validation. Multi-source Evolution: multi-source requirements driving artifact upgrades. It continuously evolved Harness capabilities such as /goal, /resume, Dynamic Workflow, Session Replay, and Desktop by converting community experience and user/developer feedback.
As of July 30, 2026, after about 16 days of fully autonomous operation, the repository had accumulated 265 commits, 127 PRs, and 151 issues, demonstrating continuous evolution in autonomous coding.
Video 1. During ten-day-scale long-horizon autonomous coding, Qwen3.8-Max autonomously built a self-evolving harness and continuously completed community-need collection, issue dispatch, code generation, validation, and self-repair.
Reproduce a paper, then surpass it
We gave the model a recent research paper, Unified Data Selection for LLM Reasoning, and asked it to reproduce the paper's experiments in code, then find a way to do better. The paper addresses a practical problem in AI training: when the amount of data far exceeds the training budget, which samples should be kept? Its answer is to select samples full of “difficult decision points”—moments during problem solving when the model is genuinely unsure which way to proceed.
The difficulty of reproducing the paper was that Qwen3.8-Max had only the paper and a batch of GPUs—no starter code and no ready-made pipeline. The data-processing scripts, training code, and evaluation program all had to be designed and written from scratch, work that even an experienced engineer would normally spend several days on.
Qwen3.8-Max worked completely independently for about five days (roughly 125 hours without interruption), wrote about 7,600 lines of code, performed more than 1,100 operations, and ran 33 rounds of GPU training. It first spent about 37 hours building the paper's complete training and evaluation pipeline from scratch, fully reproducing its six main conclusions—repeatedly fine-tuning the Qwen3-8B model on data it selected itself and validating the effect on difficult mathematics benchmarks (for example, the paper's data-selection method scored +7.7% over random selection on AIME24).
The model then went further and turned “reproduction” into “self-evolution.” Over the following roughly 88 hours, it ran a self-evolving research loop—propose a hypothesis → write code → run on a GPU → analyze → try again—across four rounds of exploration, proposing and testing 18 improvement ideas on its own. Each round's results fed the next round's hypotheses. By analyzing the problems in each approach, it eventually evolved a new method that surpassed the paper's original approach, improving its score on the competition-level mathematics benchmark AIME24 by another 2.7 points.
Expand How the improvement search unfolded step by step—4 rounds of exploration, 18 ideas Beat hundreds of human teams within 24 hours
Next, we had Qwen3.8-Max enter a real online competition—the WWW2025 multimodal conversational-intent recognition challenge, hosted on Alibaba Cloud's Tianchi platform, where 526 human teams competed. The task was to understand e-commerce customer-service conversations, including both text and screenshots, and accurately determine what the customer wanted.
Working fully independently under a 24-hour limit, Qwen3.8-Max first learned the competition rules and then built a complete solution in code. For text, it fine-tuned and ensembled multiple Chinese language models—BERT, MacBERT, and RoBERTa. For product screenshots, it fine-tuned the Qwen2.5-VL-7B vision-language model and used Chinese-CLIP as a fallback when the main model was uncertain about an image. It then combined these models into a weighted-voting system, calibrated each model's voting weight through cross-validation, and added an extra image voter to break ties. Across 45 submissions—where feedback from each submission guided the next round of fine-tuning and reweighting—its accuracy steadily rose from 0.60 to a final 0.853, beating 458 human teams (87% of the field).
Expand
Together, these three cases show what is distinctive about Qwen3.8-Max: when faced with difficult, open-ended goals, it can stay focused for days, propose its own ideas, and turn them into usable results, with no human intervention throughout.
Office assistant
Beyond coding, real work—the tedious, multi-step, tool-intensive tasks that fill nearly every profession's day—is another major path for frontier models to create economic value. Making Qwen3.8-Max genuinely capable and sufficiently stable in these workflows is our core mission.
Scaling real-world reinforcement-learning systems. By jointly scaling the number of reinforcement-learning environments and training compute, we observed continued improvement in the model's general work capabilities across mainstream harnesses (QwenWork / Claude Code / Codex / OpenClaw / Hermes). Achieving this required solving three coupled challenges:
Continuously expand real environments while decoupling them along independent dimensions: Task (single task → multi-task → multi-day task), Workspace (multiple files → hierarchical directories → complex heterogeneous directories), and Harness (different categories, versions, and skills). This allows the number of environments to grow naturally in a combinatorial way rather than relying on one-off custom integrations.
Build a unified reward system that internalizes the heterogeneous validation methods used in real tasks. The system covers execution-based checks, rubric evaluation of text and rendered visual output, and agentic checks. We unify multiple forms and modalities in one reward system, providing a consistent and reliable reward signal for all environments and eliminating the inconsistency inherent in maintaining task-specific verifiers.
Build an online data balancer that reconstructs every batch so its distributions across tasks, difficulty, workspaces, and harnesses are highly balanced, reducing gradient variance between batches and supporting stable, continuous scaling of reinforcement-learning compute.
Together, these three elements provide breadth, reliable rewards, and stability for reinforcement-learning training, turning the joint scaling of “environments × compute” into a substantial improvement in real-world work capabilities.
Figure 1. As RL training continues to scale up, Qwen3.8-Max achieves steady, consistent performance gains on dozens of in-house/public work benchmarks.
Figure 2. Qwen3.8-Max achieves comparable performance across many harnesses, including QwenWork, Claude Code, Codex, OpenClaw, and Hermes.
Testing the breadth of work capabilities across hundreds of high-value professions
The economic value of frontier models is extending into a broader range of professional settings. To test Qwen3.8-Max's ability to deliver production-grade results in real workflows, we selected frequent work scenarios from hundreds of occupations with high economic value. Here are several representative cases:
Corporate compliance lawyer — Given a corpus of hundreds of documents, Qwen3.8-Max marked 1,284 relevant clauses in one pass, completing the entire review within an hour. A document review of comparable scale would typically require a legal-assistant team working together for about a week. UI/UX designer — Qwen3.8-Max generated a high-fidelity, interactive prototype with 8 pages for the digital banking app NOVA. The unified design system worked on the first pass, with no human rework. By comparison, a similar prototype in a conventional design process requires 3–5 rounds of revisions before finalization. Restaurant brand owner — Based on more than 100 ingredient-supplier documents, Qwen3.8-Max produced a complete menu of 26 dishes in one pass. Each dish listed average calories and ingredient sources, while the ingredient cost ratio was controlled at 33.8%. Conventional menu development requires a chef and operations team to spend weeks repeatedly testing dishes, calculating costs, and adjusting the menu. Structural engineer — From a single drawing, Qwen3.8-Max recreated the seismic structural model of a 30-story office building in a browser, with the period, base shear, and inter-story drift angle available for inspection on hover. A conventional process requires an engineer to build the model manually in specialized software and typically takes more than a week. Rehabilitation therapist — Qwen3.8-Max converted a 2D paper assessment form into an interactive 3D demonstration whose anatomy diagrams can be viewed from freely selected angles and revealed layer by layer, allowing patients to understand the injury location and rehabilitation path. Such visual materials previously had to be outsourced to medical-animation teams, taking 2–4 weeks and costing thousands to tens of thousands of yuan. Sports data analyst — Qwen3.8-Max analyzed about 8,400 offensive and defensive possessions for each player into ready-to-use player tactical profiles and coaching reports, taking only several dozen minutes. A traditional analysis team would need several working days to manually perform tactical slicing, identify causes, and write the report.
Video 1. In real workflows across hundreds of high-value professions, Qwen3.8-Max substantially increases human productivity—showing the breadth of its work capabilities.
One conversation, a profitable end-to-end quantitative strategy
Drawing on its powerful Dynamic Workflows construction capability, Qwen3.8-Max can drive task planning programmatically and precisely orchestrate large-scale sub-agent systems, turning one conversation into an end-to-end, automated quantitative-research and development loop.
Depth — end-to-end ETF rotation strategy development. Starting from one short task description, Qwen3.8-Max autonomously planned a complex dynamic workflow, worked continuously for hours, and delivered a complete ETF rotation strategy. It built the data system, created baseline factors, orchestrated multiple rounds of greedy iteration, continuously analyzed backtest results, and adjusted its direction. Throughout, it acted on evidence rather than a fixed script:
When it observed the overfitting signal of metrics rising monotonically during the design period and turning downward during validation, it automatically triggered pruning and removed redundant factors round by round. When it found that multiple paths converged on the same group of core signals, it added multiple-seed-union validation to eliminate path dependence. When it judged that a three-model ensemble was less robust on a small cross-section than a fixed-direction combination, it autonomously switched to a more suitable strategy framework.
Breadth — large-scale parallel factor mining. Factor research has a huge search space, while traditional workflows are largely serial. Qwen3.8-Max parallelized the process: starting from six short descriptions covering six classic factor categories—momentum, value, quality, investment, low risk, and sentiment—it decomposed each category into 50 research directions, dispatched about 330 sub-agents, completed about 6,000 backtests, and continuously adapted the workflow while it ran. The selected factors had excess Sharpe ratios between 0.64 and 1.48, with positive daily Rank IC values between 0.010 and 0.014.
From coherent single-chain research and development to parallel exploration of a vast hypothesis space, Qwen3.8-Max uses Dynamic Workflows to encode orchestration logic as a reproducible program. It compresses quantitative research that would normally take researchers weeks or months of serial work into an automated loop that can be delivered at scale within one conversation, demonstrating broad potential for long-horizon autonomous work.
Video 2. Qwen3.8-Max may bring the expertise of quantitative researchers to everyone—showing the depth of its work capabilities.
Long-horizon tasks
When faced with extremely complex, long-cycle, multi-constraint tasks, Qwen3.8-Max demonstrates system-level autonomous planning and full-stack closed-loop adaptive learning beyond previous models. Whether designing highly precise digital chips under strong physical constraints or running a highly dynamic, competitive business simulation, the model uses an adaptive “execute–feedback–iterate” loop to deeply restructure algorithms and strategies across thousands of extremely long-horizon interactions.
Autonomously evolving chip design, with full-stack feedback optimization
Qwen3.8-Max independently completed the full chip-design pipeline of logic restructuring, multi-constraint optimization, and back-end placement and routing. The target was a GCD/RSA cryptographic hardware accelerator integrating modular exponentiation and modular multiplication. Built around a GCD datapath and control path, the design is a typical highly compact, logic-dense digital-circuit module. The model had to minimize the post-Yosys-synthesis netlist gate count (Cell Count), while ensuring bit-exact results for 4/6/8/16-bit width configurations under cocotb random verification, addressing the classic area-versus-correctness trade-off in front-end hardware design. Area was evaluated using the 16-bit (WIDTH = 16) synthesis result.
Qwen3.8-Max autonomously executed this design-optimization task in a sandbox integrating simulation (Iverilog), synthesis (Yosys), and physical-design (OpenROAD) toolchains. Its initial input was minimal: a task description, an RTL abstraction skeleton with no internal logic, and an evaluation script for functional verification and area synthesis. With no reference design and no human intervention, Qwen3.8-Max independently completed the algorithm architecture, RTL code, and subsequent rounds of iterative optimization.
In one uninterrupted autonomous run, Qwen3.8-Max went through about 500 interactions and 71 evaluations across 13 key milestones, achieving a deep end-to-end redesign. The model autonomously edited RTL code, debugged simulation, synthesized area, located redundant logic, and repeatedly restructured the datapath, progressing from fixing simulation errors and ensuring functional correctness to rewriting the algorithm. Its first working feasible design measured 8,298 gates, then it autonomously optimized it to 678 gates, the best result among all participating models. Even after hundreds of interactions, Qwen3.8-Max could still initiate major structural evolution instead of getting stuck in early local gains.
Key design milestones along the optimization trajectory (hover over each node in the evolution record to view the circuit topology and corresponding code changes):
Algorithm-level rewrite: optimizing the divider into iterative shift-subtract (8,298 → 2,010 gates, interaction 22). This was the largest single-step optimization. Qwen3.8-Max replaced the expensive 16-bit hardware modulo divider in the modular_multiplier module with a more efficient iterative Shift-Subtract architecture, removing 6,288 gates in one step and contributing more than 80% of the total reduction. Redundancy elimination and width reduction (2,010 → 1,304 gates, interactions 35–48). The model identified preconditions at the call site, safely removed the entire REDUCE stage, merged two independent overflow-reduction modules into one shared unit, optimized the output path into combinational logic, and narrowed the internal k_ff register. Register and control-state-machine pruning (1,304 → 907 gates, interactions 60–113). It removed redundant base and mod registers and the k_nz trigger, introduced an early exit for even numbers, rebuilt the comparator from the subtractor's most significant bit (MSB), and fused the GCD module's original “compare then subtract” logic into one reusable subtractor. Module fusion and logic reuse (907 → 765 gates, interactions 170–252). It broke physical module boundaries, inlined the modular multiplier into the modular-exponentiation control state machine (FSM), merged three submodules, globally shared one set of subtractors, and eliminated redundant cross-module interfaces and duplicated logic. Gate-level refinement (765 → 678 gates, interactions 443–500). It squeezed out the final local area redundancy through shared NOR gate trees, Abs-sub Splitting, and byte-to-bit selection logic.
To verify that the front-end optimization was feasible in a real physical implementation, Qwen3.8-Max sent the RTL design through a standard place-and-route flow and generated a real chip physical layout with OpenROAD using the Nangate45 process library. In the physical-layout view, standard cells are placed on the physical die plane and metal-routing layers are stacked above them, with each layer shown in a different color and connected by vertical vias. At the same physical scale, the initial and final designs differ substantially: the initial chip area was 106×106 µm², total wire length was 33,369 µm, and timing had a severe violation (Slack of -4.46 ns); the final chip area shrank to 46×46 µm², wire length fell to 4,187 µm, and timing closed successfully at 500 MHz (Slack of +0.66 ns), for an overall 81% reduction in chip area. This shows that deep front-end, architecture-level evolution can be translated seamlessly and efficiently into a physical chip with smaller area, higher speed, and better routability.
Expand
This case fully demonstrates two core capabilities of Qwen3.8-Max as a foundation model for autonomous hardware agents in long-horizon, multi-constraint tasks:
Coherent long-horizon optimization: the model can maintain a highly coherent, systematic strategy through hundreds of complex interactions and execute algorithm-level datapath restructuring rather than stopping at shallow syntax tweaks or local refactoring. Closed-loop-feedback-driven adaptive learning: without any reference implementation as a prior, the model relies entirely on the “edit–simulate–synthesize–layout” feedback loop. Every design generation must pass strict cocotb functional verification, and the OpenROAD place-and-route flow provides final closed-loop validation to ensure physical feasibility. Long-horizon operations and continuous learning
E-Commerce Bench is a 365-day long-horizon e-commerce business simulation benchmark designed to evaluate large language models' business decision-making in continuous operations. Built from real, anonymized Taobao/Tmall transaction data, it closely recreates a complex ecosystem with 12 store types, 60 product categories, nearly 600 suppliers, and 7,000 products. The model starts with ¥100,000 and operates multiple online stores at once, facing seasonal demand fluctuations, unexpected environmental events, and cash-flow pressure from a highly realistic e-commerce settlement system throughout the year. It must independently make end-to-end decisions about product selection, supply-chain negotiation, inventory management, dynamic pricing, and returns, with the ultimate goal of maximizing total cash at year end. The benchmark also tests whether it can allocate funds sensibly throughout the year, make expansionary investments, and liquidate inventory and other assets in time at the end of the cycle to avoid reducing final returns through stranded assets.
For negotiation, the benchmark introduces a game-theory-driven supplier matrix in which each supplier has independent personality traits and concession strategies, requiring the model to negotiate through multiple rounds of natural-language interaction. Qwen3.8-Max demonstrates continuous learning in negotiation: it can deeply probe the same product from the same supplier, progressively lower the purchase price and steadily increase profit, causing the radar-chart area representing negotiation efficiency to grow over time; it can also generalize this game experience effectively to similar products, while other models' negotiation efficiency generally plateaus in the middle of the run.
The model must also handle a complex market environment with hidden currents. The benchmark discreetly embeds 152 fraudulent merchants in the nearly 600-supplier matrix, covering classic fraud patterns such as “membership-fee traps,” “low-price bait,” and “goods not matching the description,” thoroughly testing risk control. At the same time, the order surge of annual promotional events is interwoven with random supply-chain crises such as typhoons and material shortages, placing extreme pressure on stocking cadence and crisis management. In this setting, Qwen3.8-Max shows outstanding forward planning: it invests the most money early in the operation, giving its subsequent asset-growth curve a faster trend; during the year-end promotion it achieves net profit above ¥100,000, nearly 2.4 times that of second-place GLM 5.2.
Qwen3.8-Max ultimately wins with total cash as high as ¥416,252 (4.16× return), 38% above second-place GLM 5.2 and 152% higher than the previous flagship Qwen3.7-Max. The result shows that Qwen3.8-Max has not only a long-horizon coherent decision-making advantage, but also the ability to learn adaptively from transaction feedback, continuing to iterate and evolve through more than 2,000 interactions rather than clinging to fixed strategies learned early on.
Multimodal agents
From everything it sees to everything it completes, Qwen3.8-Max demonstrates not merely the ability to “understand images, documents, and video,” but a form of visual productivity that runs through the entire task.
With financial reports exceeding 200 pages and complex PDFs, it can understand text, charts, and layout structure across pages, extract key conclusions from large amounts of information, and generate structured reports and deliverable web pages. With videos longer than 100 hours, it can locate specific segments and answer detailed questions, while organizing people, events, time, and scenes into a video Graph memory that continuously builds connections across long time spans and traces event evolution, relationships, and key milestones. Whether the input is hundreds of pages of documents, a complete series, or a hundred-hour livestream, information that was previously difficult to digest in one pass can be transformed into a searchable, traceable, interactive knowledge structure.
Beyond understanding, Qwen3.8-Max can complete real visual-production tasks: edit everyday footage into a Vlog, turn a problem into an immersive teaching animation, reconstruct a complete front-end project from a UI screenshot, build a Blender 3D interior-design visualization from a floor plan, and develop interactive games and applications from requirements.
More importantly, visual capability does not appear only at the input end of a task. During execution, the model continuously observes its intermediate products, checking page layout, object orientation, spatial relationships, animation effects, and interaction results. When it detects issues such as an upside-down television, a misaligned interface, or a visual effect that does not meet expectations, it can locate the deviation, re-plan, and correct it autonomously.
This means vision is no longer merely a modality an Agent uses to “understand input,” but a native feedback loop spanning planning, execution, verification, and iteration. The model generates while observing and acts while reviewing; by repeatedly “looking at the result, finding problems, and optimizing again,” it advances a task from merely completed to genuinely well done. Qwen3.8-Max is moving multimodal Agents from “understanding the world” toward continuously acting and creating in the world through vision.
In the digital world, independently finishing a complex task often requires two things at once: implementing the underlying logic in code and operating the interface directly to advance the task and observe the result. This Hybrid Agent capability—the combination of Coding and GUI operation—makes the two paths complementary: coding completes heavy work efficiently and at scale, while GUI operation reaches everything humans can see and operate, and, more importantly, brings back real feedback from a running system. It extends the visual feedback loop from the previous section, from “checking its own artifacts” to validating a real, running application.
To measure this capability, we built RecreationBench, a long-horizon “application recreation” benchmark spanning five platforms: desktop (Ubuntu, macOS, and Windows), mobile (Android), and Web. The model can observe a real running application only as a black box—it has no source code and cannot access the network—and must rely entirely on interaction and feedback to understand it, then recreate the entire application from scratch. Under this setup, Qwen3.8-Max has already demonstrated frontier-level Hybrid Agent capability, approaching the original application step by step through repeated cycles of “iterative coding—interactive feedback.”
To make this capability easier to connect to existing Agent systems, we also released Qwen-MM-Plugins. It is a Harness extension library for multimodal Agents, providing image and video processing, multimodal memory, dynamic resolution, visual tool calls, and support for specialized tasks such as video editing, Blender, and CAD across different Agent frameworks. It lets existing Agent Harnesses upgrade more naturally into multimodal-native systems.
User feedback
The most realistic judgment of Qwen3.8-Max comes from people who actually use it to do work. Leading agent-application platforms, advanced open-source algorithm teams, professional companies in law, finance, manufacturing, and other fields, emerging startups, individual developers, and academic researchers are all repeatedly giving it complex, critical, long-chain tasks.
Enterprise users use it to build and drive large-scale agent systems; office workers hand it complicated materials such as images, manuscripts, and video for one-pass processing; developers assign it demanding engineering tasks directly; and research teams use it to complete the full “literature—data—simulation” research loop. The same model is repeatedly used across different fields and workflows until people “cannot do without it,” leading to a consistent conclusion: Qwen3.8-Max can reliably take on long-chain autonomous tasks and generate high-quality, directly deliverable results in one pass.
Complete performance results Opus4.8 Fable5 GPT5.6 Sol (max) Qwen3.7-Max Qwen3.8-Max Coding Agent Terminal Bench 2.1 84.6 84.6 88.8 74.5 86.6 SWE-bench Pro 69.2 80.0 64.6 60.6 67.7 DeepSWE 1.1 59.0 70.0 73.0 21.6 56.6 NL2Repo-Bench 69.4 -- -- 47.2 55.9 FrontierSWE 70.0 88.8 -- 40.7 73.5 MLS-Bench-Lite 42.8 49.9 46.2 31.7 41.0 PaperBench 80.3 88.8 90.5 64.8 93.0 AndroidBench 69.8 84.5 74.0 56.5 75.1 QwenSWEBench 84.0 86.3 73.5 63.4 80.7 QwenQoderBench 62.7 63.1 53.8 36.8 58.4 QwenReactBench 1694 1770 1564 1538 1724 QwenSVGBench 1648 1690 1758 1499 1713 General Agent CoWorkBench 72.3 75.9 71.5 64.6 74.8 WorkSpaceBench 66.8 68.7 65.6 61.4 67.7 JobBench 48.4 57.4 45.4 31.3 53.4 SkillsBench 65.1 70.9 73.5 61.2 70.2 Agents' Last Exam (Pass / Score) 27.0 / 45.1 -- / -- 30.6 / 53.6 11.8 / 31.1 27.0 / 52.4 Automation-Bench (Pass@1) 27.2 29.1 29.7 14.2 27.3 Toolathlon Verified (Pass@1) 76.2 77.9 74.9 49.7 72.5 WideSearch 72.9 81.2 -- 75.2 81.9 HLE w/ tools 57.9 64.5 58.0 53.5 56.2 General Capabilities GPQA Diamond 92.0 92.6 94.1 92.4 92.6 HLE 45.7 53.3 47.2 41.4 43.6 IFBench 62.2 63.5 72.7 79.1 82.8 $OneMillion-Bench (expert score) 41.8 55.9 53.8 44.4 52.5 HealthBench 52.4 -- 55.3 54.5 60.2 PLawBench 69.6 70.2 72.3 58.9 73.2 PRBench-Legal 52.7 57.6 57.6 48.5 57.6 PRBench-Finance 51.9 55.8 55.5 46.8 58.3 MRCR v2 256K (8-needle) 83.2 -- 93.8 86.7 92.9 LongBench v2 69.1 -- 67.1 65.3 66.3
Fable5's results may include a fallback mechanism.
Terminal Bench 2.1: evaluated with Claude Code (avg@10), with a 5-hour timeout and max_tokens=131,072. For all other models, we report the best published score in each evaluation framework: Claude Opus 4.8 and Claude Fable 5 use Artificial Analysis's Terminus 2 (https://artificialanalysis.ai/evaluations/terminalbench-v2-1); GPT-5.6 Sol uses Codex (https://openai.com/index/previewing-gpt-5-6-sol/).
SWE-bench Pro: evaluated with the Claude Code evaluation framework, temp=1.0, top_p=0.95, and a 256K context window. Problematic tasks were corrected, and all baseline models were re-evaluated on the corrected benchmark.
DeepSWE 1.1: evaluated with the Claude Code and mini-SWE-agent frameworks, temp=1.0, top_p=0.95, and a 256K context window. We report the higher score from the two frameworks; notably, Qwen3.8-Max performs best on Claude Code.
NL2Repo-Bench: evaluated with the Claude Code framework. To prevent reward hacking, Bash commands that attempt to access specific repositories, such as pip download, pip install, and git clone, were disabled.
FrontierSWE: evaluated with the Claude Code framework. All other available MEAN@5 results come from the official FrontierSWE leaderboard (https://www.frontierswe.com), with data as of August 3, 2026. Advantage scores were recalculated from the raw scores using the official evaluation script. “--” means that no official MEAN@5 result was available as of that date.
MLS-Bench-Lite: evaluated with Claude Code, with a 5-hour timeout and max_tokens=131,072. Scores for other models come from the official leaderboard.
PaperBench: evaluated in the BasicAgent setting in Code-Dev mode, judged by Claude Opus 4.6, using the average of three runs (each run lasted up to 12 hours).
AndroidBench: evaluated on the public subset of 95 tasks; avg@3 scores are reported.
QwenSWEBench: an internal coding benchmark for evaluating software-engineering capability. Evaluated with the Claude Code framework. Reports avg@3, with an 8-hour timeout, max_tokens=32,768, temperature=1.0, and a 256K context window.
QwenQoderBench: an internal coding benchmark for evaluating the user experience in Qoder. Evaluated with the Claude Code framework. Reports avg@5, with a 6-hour timeout, max_tokens=32,768, temperature=1.0, and a 256K context window.
QwenReactBench: an internal benchmark for building React projects, using Claude Code as the evaluation framework, bilingual (English/Chinese), with seven categories; automatic rendering + multimodal judging; BT/Elo scoring.
QwenSVGBench: an internal SVG code-generation benchmark; bilingual (English/Chinese), automatic rendering + multimodal judging; BT/Elo scoring.
CoWorkBench: an internal collaboration benchmark for evaluating long-horizon tasks across computer science, finance, law, medicine, and other productivity domains.
SkillsBench: evaluated on the public SkillsBench v1.1 benchmark, covering 87 tasks, reporting the average score over three runs per task. Opus 4.8 and Fable 5 were evaluated on Claude Code; GPT-5.6 Sol on Codex; and the Qwen series on OpenCode. All results come from our own tests.
Automation-Bench: evaluated on the public subset of 600 tasks.
WideSearch: external models use the Claude Code framework, while our model uses the Qwen-Agent framework; the average item-F1 over four runs is reported.
$OneMillion-Bench: evaluated using gemini-3.1-pro-preview.
PLawBench: evaluated using gemini-3.1-pro-preview.
Blank cells (--): the score is not yet available or is not applicable.
Opus4.8 Fable5 Gemini3.1-Pro GPT5.6-Sol Qwen3.7-Plus Qwen3.8-Max Multimodal Reasoning MMMU-Pro 75.6 81.2 80.5 83.0 79.0 82.3 MathVision 87.1 / 97.1 92.7 / 98.6 87.4 / 95.7 90.8 / 97.8 90.3 / -- 95.2 / 97.7 BabyVision 28.4 / 81.2 42.5 / 90.5 55.9 / 68.3 65.5 / 88.9 64.7 / 70.4 82.0 / 91.3 HLE-VL (w/ Tools) -- -- 43.9 51.2 25.6 52.2 ZeroBench (Pass@5) 17.0 / 34.0 20.0 / 46.0 17.0 / 23.0 22.0 / 35.0 19.0 / 19.0 24.0 / 49.0 ZeroBench-Sub 31.1 37.1 36.5 46.7 41.0 48.5 LogicVista 76.7 85.7 82.6 89.7 84.3 91.9 HiPhO 69.3 78.6 85.4 86.8 84.1 90.0 PhyX 54.2 71.7 79.4 79.1 80.0 83.5 SLAKE 75.9 86.6 82.9 85.1 83.2 90.8 MedXpertQA-MM 71.7 80.0 80.7 81.5 71.0 80.4 PMC-VQA 59.2 63.2 62.5 62.3 63.4 66.2 Visual Agent & Coding OSWorld-Verified 83.4 85.0 76.2 83.2 73.3 86.1 OSWorld 2.0 20.6 / 54.8 -- / 66.1 7.8 / 30.6 -- / 62.6 2.8 / 21.5 19.4 / 46.7 ScreenSpot Pro 82.3 87.3 68.1 81.3 79.0 84.5 WebArena-Verified 67.9 71.3 64.3 69.7 55.3 66.8 AndroidWorld 75.0 88.8 70.7 77.6 81.0 85.3 MobileWorld 67.5 85.5 58.1 76.9 51.2 77.8 ClawEval-MM 73.3 / 73.8 81.2 / 77.5 50.5 / 55.2 81.2 / 78.9 57.4 / 60.1 77.2 / 74.8 Vision2Web 62.4 70.5 -- 62.1 42.1 69.0 QwenBlenderBench 62.4 69.5 23.0 68.6 41.5 69.9 Parametric CAD Bench 85.1 87.5 73.5 86.2 73.8 91.5 RecreationBench 48.0 56.1 16.2 47.6 30.2 51.7 PresentBench 80.9 79.8 55.4 82.9 65.7 79.6 Document & Office Intelligence CharXiv (RQ) 78.5 / 89.9 87.9 / 93.5 84.4 / 89.9 85.1 / 89.1 85.8 / 85.9 88.4 / 93.5 OmniDocBench 1.5 86.5 89.5 90.0 86.7 91.4 92.1 OCR-Bench-V2 (EN/ZH) 53.9 / 55.3 65.3 / 58.1 64.6 / 58.2 69.0 / 57.3 70.7 / 67.1 74.2 / 68.3 CC-OCR-Bench-V2 60.3 72.4 68.9 68.0 72.7 79.6 MTVQA-Test 48.1 41.6 54.3 52.7 51.2 56.6 MADQA 86.8 86.0 81.1 87.8 87.1 91.8 QwenVisualOffice 34.5 32.4 39.6 29.5 32.4 44.6 Real-World & Spatial Understanding RealWorldQA 76.6 85.9 83.5 83.7 86.9 88.0 ERQA 57.2 70.0 68.0 70.0 69.8 77.8 LingoQA 73.8 77.4 66.8 72.6 83.4 84.8 SURDS 62.2 79.4 64.0 63.0 77.2 77.8 Visual Perception & Grounding SimpleVQA 67.3 73.4 73.1 66.6 70.3 75.0 WorldVQA 33.9 53.5 54.0 45.1 43.9 53.2 MMStar 76.7 80.5 84.0 82.5 83.2 85.9 PerceptionBench 47.2 57.2 56.2 59.7 51.1 63.5 CountQA 41.3 63.1 72.8 68.6 77.0 82.4 RefAdv-S 61.7 68.6 71.9 69.2 73.0 80.2 Dense200 20.8 31.1 69.7 55.3 60.7 87.0 COCO 50.7 56.4 72.4 61.2 74.2 78.7 VisFactor 30.1 54.5 39.8 62.8 42.8 60.8 VLMsAreBiased 43.8 61.2 74.1 59.8 36.6 88.3 Video Intelligence & Agents VideoMME (w/ Sub.) 85.4 -- 86.7 89.5 88.0 90.4 VideoMME v2 (w/ Sub.) 49.0 52.2 66.9 71.1 59.7 68.3 VideoMMMU 75.3 81.2 85.3 85.0 85.4 88.7 MMVU 67.4 72.0 77.9 81.2 76.6 82.4 MLVU (M-Avg) 53.4 -- 84.7 87.6 87.4 90.8 TVBench 61.5 -- 73.0 83.2 78.2 81.9 LVBench 67.3 -- 75.1 78.8 76.2 81.8 LVBench (w/ Mem.) 84.3 90.1 -- 84.2 74.5 85.6 EgoLife (w/ Mem.) 78.3 82.3 -- 70.8 68.8 80.3 VideoDR (w/ Search) 65.6 77.1 -- 71.3 41.0 73.2
MathVision, BabyVision, CharXiv (RQ), and ZeroBench: scores are reported in the format “without CI / with CI.” A small number of incorrect ground-truth annotations in MathVision and CharXiv (RQ) were corrected after manual verification.
MathVision: our model was evaluated with a fixed prompt, for example, “Reason step by step and put the final answer in \boxed{}.” For other models, we report the higher score from two runs with and without the \boxed{} formatting requirement.
MMMU-Pro: results for Gemini3.1-Pro and GPT5.6-Sol come from official model reports or system cards. All other models were evaluated internally.
ClawEval-MM: scores are reported as “Pass@3 / average score.” Pass@3 measures the percentage of tasks that pass at least once in three trials; average score is the mean score over the three trials.
Vision2Web: scores are averages across the frontend, webpage, and website categories, evaluated with the Claude Code framework and gpt-5.4-2026-03-05 as the judge model.
HLE-VL (w/ Tools): evaluated with tools, including a code interpreter (CI) and search. Tool-enabled scores for Gemini3.1-Pro and GPT5.6-Sol were measured end to end through their official native tool-calling APIs.
OSWorld 2.0: scores are reported as “binary / partial.” The binary score is the percentage of tasks receiving the full task reward; the partial score aggregates partial rewards across all tasks.
ScreenSpot Pro: Opus4.8 and Fable5 scores come from official system cards. Fable5's result refers to the corresponding Mythos Preview score. All other models were evaluated internally.
WebArena-Verified: scores were reported using the official WebArena scorer within the OSWorld framework.
RecreationBench: an internal long-horizon application-recreation benchmark for evaluating Hybrid Agent capabilities across five platforms (Ubuntu, macOS, Windows, Android, and Web).
PerceptionBench: comparison-model scores come from the benchmark's official published report; our model was evaluated internally.
VideoMME (w/ Sub.) and VideoMME v2 (w/ Sub.): scores were evaluated with subtitles enabled.
QwenBlenderBench and QwenVisualOffice: both are internal benchmarks.
LVBench and EgoLife (w/ Mem.): scores were evaluated with a memory system built on Qwen-MM-Plugins, supporting fine-grained, long-horizon video memory.
VideoDR (w/ Search): scores were evaluated with the search tool enabled.
Blank cells (--): the score is not yet available or is not applicable.
Getting started with Qwen3.8
Qwen3.8-Max is now available through the Qwen AI Platform. You can integrate it seamlessly with mainstream agent frameworks and coding assistants. The model weights will also be open-sourced on Hugging Face and ModelScope next week; stay tuned.
API usage
Qwen3.8-Max officially supports reasoning_effort for adjusting reasoning depth and controlling cost:
xhigh (default): for complex tasks requiring deep analysis medium: a balance between accuracy and speed low: efficient reasoning optimized for speed and cost
In addition, preserve_thinking is enabled by default for all scenarios to provide the best out-of-the-box experience.
Qwen AI Platform
The Qwen AI Platform supports industry-standard protocols, including OpenAI-compatible Chat Completions and Responses APIs, as well as an Anthropic-compatible API interface.
python """
Environment variables:
DASHSCOPE_API_KEY: Your API Key from https://platform.qianwenai.com/home
DASHSCOPE_BASE_URL: (optional) Base URL for compatible-mode API.
- Beijing: https://dashscope.aliyuncs.com/compatible-mode/v1
- Singapore: https://dashscope-intl.aliyuncs.com/compatible-mode/v1
- US (Virginia): https://dashscope-us.aliyuncs.com/compatible-mode/v1""" from openai import OpenAI import os
api_key = os.environ.get("DASHSCOPE_API_KEY") if not api_key: raise ValueError( "DASHSCOPE_API_KEY is required. " "Set it via: export DASHSCOPE_API_KEY='your-api-key'" )
client = OpenAI( api_key=api_key, base_url=os.environ.get( "DASHSCOPE_BASE_URL", "https://dashscope.aliyuncs.com/compatible-mode/v1", ), )
messages = [{"role": "user", "content": "Write a Python function that merges two sorted linked lists."}]
completion = client.chat.completions.create( model="qwen3.8-max", messages=messages, extra_body={ "enable_thinking": True, # "preserve_thinking": True, }, reasoning_effort="xhigh", # supported levels are xhigh, medium, and low stream=True, )
reasoning_content = "" answer_content = "" is_answering = False print("\n" + "=" * 20 + "Reasoning" + "=" * 20 + "\n")
for chunk in completion: if not chunk.choices: print("\nUsage:") print(chunk.usage) continue
delta = chunk.choices[0].delta
if hasattr(delta, "reasoning_content") and delta.reasoning_content is not None:
if not is_answering:
print(delta.reasoning_content, end="", flush=True)
reasoning_content += delta.reasoning_content
if hasattr(delta, "content") and delta.content:
if not is_answering:
print("\n" + "=" * 20 + "Answer" + "=" * 20 + "\n")
is_answering = True
print(delta.content, end="", flush=True)
answer_content += delta.contentFor more information, visit the API documentation.
Code and agents
Qwen3.8-Max can be integrated seamlessly with mainstream agent frameworks and coding assistants:
Claude Code
The Qwen API supports the Anthropic API protocol and can be used directly with Claude Code:
bash npm install -g @anthropic-ai/claude-code
export ANTHROPIC_MODEL="qwen3.8-max" export ANTHROPIC_SMALL_FAST_MODEL="qwen3.8-max" export ANTHROPIC_BASE_URL=https://dashscope.aliyuncs.com/apps/anthropic
export ANTHROPIC_AUTH_TOKEN=<your_api_key>
claude
Codex
The Qwen API supports the OpenAI Responses protocol and can be used with Codex:
In ~/.codex/model-catalog.local.json
json { "models": [ { "slug": "qwen3.8-max", "display_name": "qwen3.8-max", "description": "Model Studio: Qwen3.8-Max", "default_reasoning_level": "xhigh", "supported_reasoning_levels": [ { "effort": "low", "description": "Fast responses with lighter reasoning" }, { "effort": "medium", "description": "Greater reasoning depth for complex problems" }, { "effort": "xhigh", "description": "Extra high reasoning depth for complex problems" } ], "context_window": 1000000, "effective_context_window_percent": 95, "supports_parallel_tool_calls": true, "supports_image_detail_original": true, "input_modalities": ["text", "image"], "shell_type": "default", "visibility": "list", "supported_in_api": true, "priority": 1, "base_instructions": "", "support_verbosity": false, "supports_reasoning_summaries": false, "experimental_supported_tools": [], "truncation_policy": { "mode": "bytes", "limit": 10000 } } ] }
In ~/.codex/config.toml
toml model_catalog_json = "~/.codex/model-catalog.local.json"
model_provider = "ModelStudio" model = "qwen3.8-max"
[model_providers.ModelStudio] name = "Model Studio" base_url = "https://dashscope.aliyuncs.com/compatible-mode/v1" env_key = "OPENAI_API_KEY" wire_api = "responses" bash npm install -g @openai/codex
export OPENAI_API_KEY=<your_api_key>
codex
Qoder CLI
Qoder evolves together with Qwen, focusing on agentic coding:
bash curl -fsSL https://qoder.com/install | bash
qoder
Qwen Code
Qwen Code is deeply optimized for the Qwen series:
bash npm install -g @qwen-code/qwen-code@latest
qwen
OpenClaw
Connect OpenClaw through the Qwen AI Platform:
bash curl -fsSL https://molt.bot/install.sh | bash
export DASHSCOPE_API_KEY=<your_api_key>
openclaw dashboard
Edit ~/.openclaw/openclaw.json to configure it:
json { "models": { "mode": "merge", "providers": { "bailian": { "baseUrl": "https://dashscope.aliyuncs.com/compatible-mode/v1", "apiKey": "DASHSCOPE_API_KEY", "api": "openai-completions", "models": [ { "id": "qwen3.8-max", "name": "qwen3.8-max", "reasoning": true, "input": ["text", "image"], "contextWindow": 1000000, "maxTokens": 65536 } ] } } }, "agents": { "defaults": { "model": { "primary": "bailian/qwen3.8-max" } } } }
Summary
Qwen3.8-Max is our most powerful model to date and the first Max-scale model with open-source weights. Its parameter scale has expanded to 2.4 trillion, bringing comprehensive improvements in coding, real office work, long-horizon tasks, and multimodal agents. It can complete complex, open-ended goals end to end with very little human intervention and deliver trustworthy results. The model weights will be open-sourced next week; stay tuned. We welcome community feedback and look forward to seeing what everyone creates.
Citation bibtex @misc{qwen38, title = {Qwen3.8-Max: A New Bar for Coding and Cowork}, url = {https://qwen.ai/blog?id=qwen3.8}, author = {{Qwen Team}}, month = {August}, year = {2026} } Use Qwen Studio Web iOS Android macOS Windows Qwen Studio Qwen Studio overview Download API Platform Flagship models Platform overview API Platform Qwen Cloud Research Latest developments Research index GitHub Terms and policies Terms of Use Privacy Policy Usage Policy Cookie notice Training Data Disclosure Summary Qwen © 2026 Manage cookies Powered by Alibaba Cloud
The “Source article” above is the visible body extracted from the Tabbit international page and includes the opening navigation text and closing footer text; it was not accessed through the China edition or domestic sites. The internal benchmarks, cases, and parameters in the source should be checked against the official page again before publication.
Qwen3.8 Max