# Qwen3.8-Max

> Source: https://aiwiki.ai/wiki/qwen3_8_max
> Updated: 2026-09-05
> Fact-checked: 2026-09-05
> Categories: AI Agents, AI Code Generation, Chinese AI, Large Language Models
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Qwen3.8-Max." aiwiki.ai, 5 Sept 2026. https://aiwiki.ai/wiki/qwen3_8_max
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**Qwen3.8-Max** is a 2.4-trillion-parameter mixture-of-experts [large language model](https://aiwiki.ai/wiki/large_language_model) developed by Alibaba's [Qwen](https://aiwiki.ai/wiki/qwen) team and released on August 3, 2026 under the title "Qwen3.8-Max: A New Bar for Coding and Cowork".[1] It is the production flagship of the [Qwen3.8](https://aiwiki.ai/wiki/qwen3_8) generation, which had been available since July 19, 2026 only as the hosted `qwen3.8-max-preview` service.[5] Alibaba positioned it as its most capable Qwen model and, in a break with the closed-weight Qwen-Max line that includes [Qwen3-Max](https://aiwiki.ai/wiki/qwen3_max) and [Qwen3.7-Max](https://aiwiki.ai/wiki/qwen3_7_max), announced that it would become the first Max-class model released with [open weights](https://aiwiki.ai/wiki/open_weights).[1][2] Qwen published the text-only `Qwen/Qwen3.8-2.4T-A95B` checkpoint on August 12 under the custom Qwen3.8-Max License; unlike the hosted service, the open checkpoint requires thinking mode and does not accept images or video.[14][15][16] On September 2, Qwen introduced the dated hosted snapshot `qwen3.8-max-0902`, described as a further post-trained update for coding and professional work.[17][18]

## Key facts

| Item | Detail |
| --- | --- |
| Developer | Qwen team, Alibaba |
| Released | August 3, 2026; open checkpoint August 12; 0902 hosted snapshot September 2 [1][5][14][17] |
| Parameters | 2.4 trillion total, 95 billion active per token [1] |
| Architecture | Mixture-of-experts, "built upon the architectural foundation of Qwen3.5" [1][4] |
| Hosted context window | 1M tokens (max input 991K, max output 131K, max reasoning 262K) [4][18] |
| Hosted modalities | Text, image, and video input; text output [4][18] |
| API pricing | $2 per 1M input tokens, $6 per 1M output tokens, $0.25 per 1M implicitly cached [2][4] |
| Access | QwenCloud API (`qwen3.8-max`; dated snapshot `qwen3.8-max-0902`), Qwen Studio chat [1][2][18] |
| Weights | `Qwen/Qwen3.8-2.4T-A95B`, released August 12; text-only, thinking-required, Qwen3.8-Max License [14][15][16] |

## Release

Alibaba introduced the Qwen3.8 generation on July 19, 2026 as `qwen3.8-max-preview`, available through Token Plan, Qoder, and QoderWork, describing a continuously evolving 2.4-trillion-parameter model that the company claimed was "second only to Fable 5" and would go open weight.[5] The production release followed on August 3, 2026, announced in a Qwen team blog post published at 10:00 Beijing time and an accompanying thread from the official Qwen account.[1][2] Alibaba Group's corporate account called it "the largest and most capable flagship model to date" with 2.4 trillion parameters and a 1 million-token context window.[3]

The release landed within hours of another Chinese frontier release: MiniMax announced its open-weight [MiniMax-H3](https://aiwiki.ai/wiki/minimax_h3) less than an hour after Qwen's first post the same morning.[9]

The blog post states that "this also marks the first time we will open-source the weights of a Qwen-Max-class model", while the launch thread added that Qwen3.8-27B, a smaller model, "is also going open-weights".[1][2] Analyst Ben Thompson, in commentary relayed by Simon Willison when the plan was first announced in July, theorized that the decision, a reversal of Alibaba's choice not to release Qwen3.7-Max weights in May 2026, may have been influenced by Xi Jinping's public call to "encourage open source, openness, collaboration and sharing".[8] Qwen released `Qwen3.8-2.4T-A95B` on Hugging Face and ModelScope on August 12.[14] The checkpoint has 2.4 trillion total parameters and 95 billion active parameters, native context of 262,144 tokens that can be extended to 1,010,000, and a custom Qwen3.8-Max License.[15][16] It is text-only and requires thinking mode. Qwen's hosted Max service is based on the checkpoint but adds vision input, non-thinking support, a 1 million-token context by default, and built-in tools.[15]

### Qwen3.8-Max-0902 update

Qwen announced `Qwen3.8-Max-0902` on September 2, 2026 as a dated update to the hosted model.[17] The company said it applied further post-training for coding and "Cowork" and described stronger performance on complex enterprise work, scientific research, and long-horizon workflows.[17] These are Qwen's descriptions of the update, not results from an independent evaluation.

QwenCloud identifies the snapshot as `qwen3.8-max-0902`, with `qwen3.8-max-2026-09-02` as an alias.[18] Its catalog keeps the hosted model's text, image, and video inputs; text output; 1 million-token context; and maximum reasoning budget of 262,000 tokens. The listed international price also remains $2 per million input tokens and $6 per million output tokens.[18]

## Positioning: coding and "Cowork"

The "Cowork" in the release title is not a separate product. Alibaba uses the word for the model's ability to carry out real professional work autonomously inside agent harnesses, the subject of the blog post's "Work" section, and measures it with CoWorkBench, an in-house benchmark the company describes as covering "long-horizon tasks across computer science, finance, law, medical, and other productivity domains".[1] Alibaba says it trained this capability by jointly scaling reinforcement-learning environments and compute across several harnesses, naming QwenWork, Claude Code, Codex, OpenClaw, and Hermes, with a unified reward system spanning execution checks, rubric-based judging, and agentic inspection.[1] Alibaba Cloud's launch messaging framed the model as an "always-on workmate" for coding, reasoning, research, writing, and AI agents.[2]

## Vendor demonstrations

All of the following results are Alibaba's own reported outcomes, not independent evaluations.[1]

In coding, Alibaba's headline demonstration is a "10+ day" autonomous run in which the model built oh-my-cli, a command-line agent project, from an empty folder: requirements entered GitHub issues, agents claimed and executed them through a state machine, and changes passed end-to-end tests before merging. Alibaba says that as of July 30, 2026, after roughly 16 days of fully autonomous operation, the repository held 265 commits, 127 pull requests, and 151 issues; the project trace is public on GitHub.[1][10] A second demonstration had the model reproduce a research paper on data selection for LLM reasoning from nothing but the paper and GPUs, then improve on it: about five days of continuous work, roughly 7,600 lines of code, 33 rounds of GPU training, and a final method Alibaba says beats the paper's own approach by 2.7 points on AIME24.[1] In a third, the model entered a live Tianchi data-science competition under a 24-hour limit and, across 45 submissions, reached an accuracy of 0.853, which Alibaba says beat 458 of 526 human teams.[1]

For long-horizon tasks, Alibaba reports that the model autonomously optimized a cryptographic hardware accelerator design over roughly 500 turns, cutting the synthesized design from 8,298 to 678 gates and passing place-and-route timing closure at 500 MHz in an OpenROAD flow.[1] On E-Commerce Bench, a 365-day simulated store-operation benchmark built on Taobao and Tmall transaction data, Alibaba says the model turned 100,000 yuan of starting capital into a top-ranked 416,252 yuan, 38 percent ahead of second-place GLM 5.2 and a 152 percent improvement over Qwen3.7-Max.[1]

On the multimodal side, Alibaba says the model can work across financial reports and PDFs longer than 200 pages and videos longer than 100 hours, organizing people, events, and scenes into a "video memory graph", and can pair coding with GUI operation as a "Hybrid Agent" that verifies its own output against a running application. Alongside the model, the company introduced RecreationBench, an internal benchmark in which a model must rebuild a running application it can only observe as a black box, and Qwen-MM-Plugins, a harness extension library for multimodal agents.[1]

## Benchmarks

Alibaba published a large self-reported comparison table against Anthropic's Claude Opus 4.8 and Claude Fable 5, OpenAI's GPT-5.6 Sol, and its own Qwen3.7-Max (with Qwen3.7-Plus as the baseline on multimodal rows). Selected scores:[1]

| Benchmark (Alibaba-reported) | Qwen3.8-Max | Qwen3.7-Max | Claude Fable 5 | GPT-5.6 Sol |
| --- | --- | --- | --- | --- |
| Terminal Bench 2.1 | 86.6 | 74.5 | 84.6 | 88.8 |
| SWE-bench Pro | 67.7 | 60.6 | 80.0 | 64.6 |
| FrontierSWE | 73.5 | 40.7 | 88.8 | not listed |
| PaperBench | 93.0 | 64.8 | 88.8 | 90.5 |
| CoWorkBench | 74.8 | 64.6 | 75.9 | 71.5 |
| GPQA Diamond | 92.6 | 92.4 | 92.6 | 94.1 |
| Humanity's Last Exam | 43.6 | 41.4 | 53.3 | 47.2 |
| IFBench | 82.8 | 79.1 | 63.5 | 72.7 |

Alibaba's own table therefore does not show a uniform lead: Qwen3.8-Max posts the top listed score on PaperBench, IFBench, and many multimodal rows such as OSWorld-Verified (86.1) and OmniDocBench 1.5 (92.1), but trails Claude Fable 5 on several coding and reasoning benchmarks, including SWE-bench Pro, FrontierSWE, and Humanity's Last Exam.[1] The table's footnotes matter for interpretation: many rows were evaluated with the Claude Code harness, several benchmarks (PaperBench, CoWorkBench, QwenSWEBench, RecreationBench, and others) are Alibaba's own in-house constructions or were "corrected" versions of public sets, and one footnote states that "Fable5 results may involve fallbacks".[1]

Independently, [Artificial Analysis](https://aiwiki.ai/wiki/artificial_analysis) scored Qwen3.8-Max at 53 on its Intelligence Index (v4.1) on release day, which it described as well above the median of 32 for reasoning models in a similar price tier. The evaluator also flagged extreme verbosity, with the model generating 150 million tokens during the index run against a median of 63 million, and measured output speed of 46.5 tokens per second on Alibaba's API, below the tier median of 72.0.[7] Because Artificial Analysis has revised its index methodology between releases, this score is not directly comparable to the 56.6 that Qwen3.7-Max recorded on an earlier index version.

### Code Arena: WebDev result

On September 2, Qwen said the 0902 snapshot had moved from 1669 to 1691 on Arena's Code Arena: WebDev leaderboard and taken the nominal number-one position.[19] The live Arena leaderboard corroborated the score and rank when accessed that day. Its displayed August 31 snapshot listed `qwen3.8-max-0902` first at 1691, with a preliminary interval of +19/-19 from 1,389 votes.[20] The earlier `qwen3.8-max` remained separately listed at 1669.

| Nominal rank | Model | Arena score | Votes | Displayed rank spread |
| --- | --- | --- | --- | --- |
| 1 | qwen3.8-max-0902 | 1691 +19/-19 (preliminary) | 1,389 | 1-4 |
| 2 | claude-opus-5-max | 1688 +8/-8 | 10,334 | 1-4 |
| 3 | kimi-k3-max | 1674 +11/-11 | 4,544 | 1-5 |
| 4 | qwen3.8-max | 1669 +12/-12 (preliminary) | 3,219 | 1-5 |

The three-point nominal lead over Claude Opus 5 Max was smaller than either model's displayed interval, and Arena gave Qwen3.8-Max-0902 a rank spread of 1-4.[20] The result therefore supports a dated, preliminary first-place description, not a statistically decisive lead.

Code Arena: WebDev measures human preference on generated web applications rather than general software correctness. Users submit a prompt, compare two anonymous apps, and vote for the better result; Arena aggregates the pairwise outcomes with a Bradley-Terry model.[21] The original WebDev Arena implementation used a constrained frontend environment based on React, TypeScript, Tailwind CSS, and Next.js, and Arena's current board describes its scope as front-end web development tasks that can include multistep reasoning and tool use.[20][21]

### RSI-Exam result

On September 4, 2026, Huaxiu Yao, whose AIMING Lab runs the [RSI-Exam](https://aiwiki.ai/wiki/rsi_exam) benchmark of long-horizon executable research tasks, posted that Qwen3.8 Max-0902 had moved "from 0.322 to 0.392, a nearly 22% improvement" on the RSI-Exam 0.1 leaderboard.[22] Qwen Developers quoted the post: "We did comprehensive training on coding and cowork for Qwen3.8-Max-0902, with complex, long-horizon tasks in mind. Glad to see that work generalizes to a 22% improvement on RSI-Exam."[23] As of September 5, 2026, the full 88-task board listed Qwen3.8 Max-0902 fourth at 0.392 and the original Qwen3.8 Max ninth at 0.322, behind Claude Opus 5 (0.464), GPT-5.6-sol (0.433), and GLM 5.3 (0.403); every score sat below the benchmark's 0.6 frontier-calibrated reference.[24] The two Qwen rows were not run the same way: the 0.322 result used the qwen coder harness at xhigh reasoning effort, while the 0.392 result used Claude Code at xhigh.[24] RSI-Exam's authors state that tool use, context management, and stopping behavior are part of the evaluated system and that "leaderboard differences should not be attributed to the underlying model alone," so the 22% figure combines the harness change with the snapshot update and is not a like-for-like measurement of the 0902 update.[25]

## Availability, API, and pricing

Qwen3.8-Max is served through QwenCloud under the model identifier `qwen3.8-max`, with OpenAI-compatible chat-completions and responses endpoints, an Anthropic-compatible API, and endpoints in Beijing, Singapore, and the United States (Virginia); it is also available to chat users in Qwen Studio.[1][2] QwenCloud lists the September 2 snapshot as `qwen3.8-max-0902`, with `qwen3.8-max-2026-09-02` as an alias.[18] The hosted model has a 1M-token context window with a maximum input of 991K tokens, maximum output of 131K tokens, maximum reasoning budget of 262K tokens, and text, image, and video input.[4][18]

The API exposes a `reasoning_effort` parameter with levels `low`, `medium`, and `xhigh` (the default), and enables `preserve_thinking` by default so prior reasoning content carries across turns.[1] Alibaba documents integrations with Claude Code, Codex, Qoder CLI, Qwen Code, and OpenClaw.[1] Pricing for the 0902 snapshot is $2.00 per million input tokens and $6.00 per million output tokens, with implicit cache hits billed at $0.25 per million; QwenCloud also lists explicit cache creation at $2.50 and explicit cache reads at $0.17 per million tokens.[18]

## Reception

Coverage centered on the scale claim and the open-weights turn. The Decoder reported the release as Alibaba positioning a 2.4-trillion-parameter model, with 95 billion active parameters per query, for long-horizon autonomous work, and noted that Qwen3.8-Max is the first Qwen-Max-class model whose weights were planned for publication.[6] Simon Willison, who had tested the preview, highlighted the model's reasoning traces during his informal SVG-drawing test and observed that at 2.4 trillion parameters the model is nearly as large as the 2.8-trillion-parameter Kimi K3.[8] Qwen later published the weights and license, but the open checkpoint is text-only while the hosted service retains additional multimodal and non-thinking features.[14][15][16] The launch demonstrations above still rest mainly on Alibaba's reporting rather than independent replication.[1][6]

On August 3, 2026, the Qwen account also posted a screenshot of the [LMArena](https://aiwiki.ai/wiki/lmarena_org) Text Arena leaderboard with the caption "Alibaba hits #2 on the Text Arena": in the leaderboard's lab-ranking view, dated August 1, 2026, Alibaba placed second among model providers behind Anthropic on the strength of Qwen3.8-Max, whose model-level rank in the same screenshot was fifth.[12] As of August 5, 2026, the live text leaderboard showed qwen3.8-max in fifth place among individual models, with a preliminary Arena score of 1496 from 3,327 votes, the highest placement of any non-Anthropic model: behind Claude Fable 5 (1509) and three Claude Opus variants, and ahead of Google's Gemini 3 Pro and OpenAI's GPT-5.6 Sol.[13]

## See also

- [Qwen3.8](https://aiwiki.ai/wiki/qwen3_8)
- [Qwen](https://aiwiki.ai/wiki/qwen)
- [Qwen3.7-Max](https://aiwiki.ai/wiki/qwen3_7_max)
- [Qwen3-Max](https://aiwiki.ai/wiki/qwen3_max)
- [Qwen3-Coder](https://aiwiki.ai/wiki/qwen_3_coder)
- [MiniMax-H3](https://aiwiki.ai/wiki/minimax_h3)

## References

1. Qwen Team. "Qwen3.8-Max: A New Bar for Coding and Cowork." August 3, 2026. https://qwen.ai/blog?id=qwen3.8
2. Qwen (@Alibaba_Qwen). Qwen3.8-Max release thread (pricing, open-weights plan, Qwen3.8-27B). X, August 3, 2026. https://x.com/Alibaba_Qwen/status/2084100707423289643
3. Alibaba Group (@AlibabaGroup). Qwen3.8-Max announcement post. X, August 3, 2026. https://x.com/AlibabaGroup/status/2084102898796105964
4. QwenCloud. "Qwen3.8-Max" model page. Accessed August 5, 2026. https://www.qwencloud.com/models/qwen3.8-max
5. Qwen (@Alibaba_Qwen). Qwen3.8-Max-Preview launch post. X, July 19, 2026. https://x.com/Alibaba_Qwen/status/2078759124914098291
6. The Decoder. "Alibaba's open-weight Qwen3.8-Max takes on long-horizon AI tasks with 2.4 trillion parameters." August 3, 2026. https://the-decoder.com/alibabas-open-weight-qwen3-8-max-takes-on-long-horizon-ai-tasks-with-2-4-trillion-parameters/
7. Artificial Analysis. "Qwen3.8 Max - Intelligence, Performance & Price Analysis." Accessed via Internet Archive snapshot of August 3, 2026. https://web.archive.org/web/20260803215115/https://artificialanalysis.ai/models/qwen3-8-max
8. Simon Willison. "Who's Afraid of Chinese Models?" July 20, 2026. https://simonwillison.net/2026/Jul/20/afraid-of-chinese-models/
9. MiniMax (@MiniMax_AI). "MiniMax-H3 Is Now Publicly Available." X, August 3, 2026. https://x.com/MiniMax_AI/status/2084106804032872591
10. qwen-code-dev-bot. "oh-my-cli" repository. GitHub. Accessed August 5, 2026. https://github.com/qwen-code-dev-bot/oh-my-cli
11. Qwen. Hugging Face organization page. Accessed August 5, 2026. https://huggingface.co/Qwen
12. Qwen (@Alibaba_Qwen). "Alibaba hits #2 on the Text Arena." X, August 3, 2026. https://x.com/Alibaba_Qwen/status/2084111492182659552
13. LMArena. Text Arena leaderboard. Accessed August 5, 2026. https://lmarena.ai/leaderboard/text
14. QwenLM. "Qwen3.8" repository (August 12 open-checkpoint release notice). GitHub. Accessed September 2, 2026. https://github.com/QwenLM/Qwen3.8
15. Qwen. "Qwen3.8-2.4T-A95B" model card. Hugging Face. Accessed September 2, 2026. https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B
16. Qwen. "Qwen3.8-Max License." Hugging Face. Accessed September 2, 2026. https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B/blob/main/LICENSE
17. Qwen (@Alibaba_Qwen). Qwen3.8-Max-0902 release post. X, September 2, 2026. https://x.com/Alibaba_Qwen/status/2094968708288680276
18. QwenCloud. "Qwen3.8-Max-0902" model page. Accessed September 2, 2026. https://www.qwencloud.com/models/qwen3.8-max-0902
19. Qwen (@Alibaba_Qwen). "#1 on CodeArena: WebDev leaderboard." X, September 2, 2026. https://x.com/Alibaba_Qwen/status/2094976556494209206
20. Arena.ai. "Code Arena: WebDev, Overall" live leaderboard. Accessed September 2, 2026. https://arena.ai/leaderboard/code/webdev
21. Arena Team. "WebDev Arena: A Live LLM Leaderboard for Web App Development." Published March 10, 2025; updated June 13, 2025. https://arena.ai/blog/webdev-arena
22. Huaxiu Yao (@HuaxiuYaoML). Post on Qwen3.8 Max-0902's RSI-Exam 0.1 score. X, September 4, 2026. https://x.com/HuaxiuYaoML/status/2095861167939805252
23. Qwen Developers (@QwenDevs). Post on Qwen3.8-Max-0902 training and its RSI-Exam improvement. X, September 4, 2026. https://x.com/QwenDevs/status/2095888382064988439
24. RSI-Exam Team. "RSI-Exam: Benchmarking Recursive Self-Improvement through Executable Research" (leaderboard, Full 88 board). Accessed September 5, 2026. https://rsi-exam.ai/
25. RSI-Exam Team. "RSI-Exam: Benchmarking Recursive Self-Improvement through Executable Research" (blog post, "Limitations and next steps"). Accessed September 5, 2026. https://rsi-exam.ai/blog.html