The Qwen4-Max launch page reports a score of 52.69 on Work360, an industry benchmark built by Qwen, and claims lower task-token use and average cost per task than Qwen3.8-Max. Other launch charts combine official retests with scores cited from third-party leaderboards, so they should not be treated as results from one evaluation suite or as independently verified findings.
Tasks this can help assess: Qwen's reported benchmark results for professional work, long-horizon agents, coding, and multimodal tasks.
Tasks these results cannot be generalized to: They do not establish success rates across all real-world workflows, and scores from different leaderboards should not be compared directly.
Applicable model versions: Qwen4-Max; some comparisons use Qwen3.8-Max or Qwen3.8-Max-0902.
Test environment or client: The launch page does not fully disclose a unified runtime environment, tool configuration, or review process. The model is available on the Qwen AI platform, through the API, and in Qwen Office, Qoder CN, the Qwen app, and Qwen Studio.
Reasoning levels and parameters: Qwen reports token-efficiency comparisons at High and Xhigh reasoning levels for Work360 and Terminal-Bench 4.0; other parameters are not specified.
Qwen defines Work360 as a comprehensive framework covering industrial and professional work, with a taxonomy of 26 top-level domains, 182 second-level subdomains, and 721 third-level subdomains. The launch page does not provide the sample size, task-sampling method, scoring rubric, tool-calling configuration, or error intervals.
The launch charts list multiple benchmarks for coding, office work, agents, and professional tasks. A note at the bottom of the page says some figures are citations from third-party leaderboards: APEX-Agents v1.1 is cited from Mercor; AA-Briefcase v1.1 and AutomationBench-AA from Artificial Analysis; CodeArena and AgentArena from Arena; HLAB-Vals-1.1 from Vals Index; and Agents' Last Exam v1.1 from its leaderboard. JobBench was computed using an updated dataset and evaluation code. The figures therefore come from different sources and evaluation processes.
Work360 (Qwen-built benchmark): Qwen4-Max scored 52.69, compared with 47.74 for Qwen3.8-Max. Qwen says the new model covers more industry tasks. The evaluation set and scoring implementation have not been disclosed in enough detail for independent reproduction.
Work360 token use and cost (official comparison): At Xhigh and High reasoning levels, Qwen4-Max used about 24% and 35% fewer tokens, respectively. Qwen reports that average cost per Work360 task fell from RMB 26.58 to RMB 11.16, a reduction of about 58%. The page does not specify the number of tasks, the pricing interval, or the exact task distribution across reasoning levels.
Terminal-Bench 4.0: The launch chart shows a score of 48.5 for Qwen4-Max. Qwen says token use fell by about 48% at Xhigh and 49% at High. The chart does not provide the complete runtime configuration in the main text, so the score alone cannot establish equivalence with results from other environments.
JobBench (official retest): The launch chart shows a score of 68.5. Qwen notes that it used an updated dataset and evaluation code; this result should be distinguished from results on the previous version.
Third-party leaderboard citations: The chart shows 72.1 on APEX-Agents v1.1, 1686 Elo on AA-Briefcase v1.1, 72.5 on AutomationBench-AA, 1750 Elo on CodeArena, 21.3 on HLAB-Vals-1.1, and 46.7 on Agents' Last Exam v1.1. Qwen says these results come from the corresponding third-party leaderboards or evaluators. They are figures cited on the launch page, not equivalent to retests by the Qwen team in one unified local environment.
Internal base-model evaluations: Qwen says Qwen4-Max-Base had the lowest 10×BPB among five evaluated models in its internal evaluations of knowledge, code, question answering, and STEM. Compared with Qwen3.8-Max-Base, BPB fell by about 5–13% across domains, while active parameters per token fell by about 25%. An official footnote cautions that GLM-5.3 and Kimi-K3.1-Pro, among the evaluated models, are post-trained models; comparisons with base models are affected by training stage.
| Item | Launch-page data | Source/method |
|---|---|---|
| Work360 | 52.69; comparison: 47.74 | Qwen-built industry benchmark |
| Average Work360 cost per task | RMB 26.58 → RMB 11.16 (−58%) | Official report; task count and pricing interval not specified |
| Work360 token use | Xhigh −24%; High −35% | Official report |
| Terminal-Bench 4.0 | 48.5; token use: Xhigh −48%, High −49% | Official launch chart and report |
| JobBench | 68.5 | Official retest using an updated dataset and evaluation code |
| APEX-Agents v1.1 | 72.1 | Cited from the Mercor leaderboard |
| AA-Briefcase v1.1 / AutomationBench-AA | 1686 Elo / 72.5 | Cited from Artificial Analysis |
| CodeArena / HLAB-Vals-1.1 / Agents' Last Exam v1.1 | 1750 Elo / 21.3 / 46.7 | Cited from Arena, Vals Index, and Agents' Last Exam, respectively |
The official material supports two limited conclusions: Qwen4-Max scored higher than its previous-generation comparator on Qwen's Work360 benchmark, and Qwen reports lower token use and cost per task at several reasoning levels. It does not disclose enough Work360 samples, scoring details, or run traces to verify reproducibility. Third-party results in the launch charts should be interpreted according to each leaderboard's version, scale, and evaluation method; Elo ratings, percentage scores, and results on different task sets cannot be ranked directly against one another.
The launch page describes the model as released and available on the Qwen platform and the listed products. Its specifications are 2.4T backbone parameters, 71B active parameters per token, and an additional 328B N-gram Embedding parameters. It has native vision capabilities and a 1M-token context window. The launch page does not specify an API model ID or provide the full inference parameters or benchmark harness.
This note records a review of the official launch page and its charts on 2026-11-17. The page does not provide the complete Work360 data, task samples, or an executable evaluation package. Third-party leaderboards must be checked on their respective pages against the relevant versions. Third-party scores in the charts should retain their source attribution and should not be recast as results from independent experiments.
Qwen4-Max