Quick Answer
Alibaba released Qwen3.8-Max on August 3, 2026, a 2.4-trillion-parameter, multimodal mixture-of-experts model available now via API at $2 per million input tokens and $6 per million output tokens. Open weights are scheduled for Hugging Face during the week of August 10. US users and businesses should keep sensitive data out of the hosted API because prompts are processed under Chinese data jurisdiction.
Key Takeaways
- Qwen3.8-Max has 2.4 trillion total parameters but activates only 95 billion per token, using a sparse Mixture-of-Experts architecture that keeps per-request compute lower than the headline number suggests.
- The model supports text, image, and video input with a 1-million-token context window and up to 131,072 output tokens per response.
- Alibaba’s self-reported benchmarks place Qwen3.8-Max at 86.1 on OSWorld-Verified, ahead of GPT-5.6 Sol Max and Claude Fable 5 on that specific test, but independent verification across all benchmarks remains limited as of August 8, 2026.
- Open weights for Qwen3.8-Max and a smaller Qwen3.8-27B are scheduled for release on Hugging Face and ModelScope during the week of August 10, 2026, the first time Alibaba has open-sourced a Max-class model.
- Sending work, financial, health, or otherwise sensitive data through the hosted QwenCloud API carries a data-jurisdiction risk because Alibaba operates under Chinese law, including the Cybersecurity Law and the Data Security Law.
What is Qwen3.8-Max and when did it launch?
Qwen3.8-Max is Alibaba’s most capable AI model to date, released on August 3, 2026, following a preview at the World AI Conference in Shanghai on July 19. Qwen3.8-Max is the largest model in Alibaba’s Qwen family, featuring 2.4 trillion parameters and a context window of up to 1 million tokens. The model has roughly seven times more parameters than the Qwen3.5 model that Alibaba released in February of this year.
Despite its total size of 2.4 trillion parameters, Qwen3.8-Max activates only 95 billion parameters at inference time through a Sparse Mixture-of-Experts architecture. This architectural choice matters practically because the actual compute load during a query is much closer to a mid-sized dense model than the 2.4-trillion headline implies, which directly affects API speed and cost. This release also marks the first time Alibaba will open-source the weights of a Qwen Max-class model, built on the architectural foundation of Qwen 3.5.
It is also worth clarifying a common source of confusion: Qwen3.8-Max is not the same as Qwen3-8B. Qwen3.8 Max is not Qwen3-8B. Qwen3-8B scores describe an older eight-billion-parameter checkpoint and must not be used as Qwen3.8 evidence.
What can Qwen3.8-Max actually do?
Qwen3.8-Max handles text, image, and video input in a single API endpoint, and returns text responses of up to 131,072 tokens. Qwen3.8-Max is the flagship model in Alibaba’s Qwen3.8 series, intended for complex reasoning, visual understanding, coding, and agentic workflows. The model supports prompts with up to 1 million tokens of data, which, according to Alibaba, enables it to analyze more than 200 pages of text or about 100 hours of footage per request. These figures are Alibaba’s own claims and have not been independently verified at scale.
Built on the Qwen3.5 architecture, the model is aimed at coding, research, professional work, and multimodal tasks. Alibaba highlighted autonomous coding as a centerpiece of the launch. In one company-run test, Qwen3.8-Max spent about 16 days building and maintaining the oh-my-cli project, producing 265 commits, 127 pull requests, and 151 issues. That demonstration is a vendor showcase rather than an independent reproduction, and the practical performance of agentic coding runs will vary by task complexity and tooling setup.
Thinking is always enabled, with low, high, and xhigh reasoning settings, with xhigh as the documented default. For most users, that default setting provides the deepest analysis, though it also consumes more tokens per response and raises the per-query cost accordingly.
How does Qwen3.8-Max compare to GPT-5.6 and Claude Fable 5?
Qwen3.8-Max is competitive with the current top tier of frontier models on several benchmarks, but it does not lead across all task types. The comparison requires care because most scores available as of August 8, 2026 are Alibaba’s own self-reported figures, and independent verification is still limited.
| Benchmark | Qwen3.8-Max | Claude Fable 5 | GPT-5.6 Sol Max | Notes |
|---|---|---|---|---|
| OSWorld-Verified (agentic computer use) | 86.1 | 85.0 | 83.2 | Alibaba self-reported |
| PaperBench (research reproduction) | 93.0 | N/A | N/A | Alibaba self-reported |
| SWE-bench Pro (software engineering) | 67.7 | Leads | N/A | Alibaba self-reported; Fable 5 leads |
| Text Arena (crowdsourced ranking) | 5th globally | Top 3 | N/A | Snapshot; ranks shift daily |
| Vision Arena (crowdsourced ranking) | 2nd globally | 1st | N/A | Alibaba-reported snapshot |
Alibaba’s headline claim is that Qwen3.8-Max scores 86.1 on OSWorld-Verified, ahead of GPT-5.6 Sol Max at 83.2, Claude Fable 5 at 85.0, and Gemini 3.1 Pro at 76.2. At the same time, the picture is not uniformly favorable. On several other benchmarks, Claude Fable 5 leads on SWE-Pro, QwenReactBench, CoWorkBench, JobBench, and MobileWorld, while GPT-5.6 Sol leads on TerminalBench-2.1, FrontierSWE, and Agents’ Last Exam.
The most important caveat is that most of the scores driving the coverage are not independently replicated. As of August 4, Alibaba has published no benchmark table, no model card, and no methodology behind its “second only to Fable 5” claim. The practical takeaway is that Qwen3.8-Max is competitive with, and in several categories ahead of, the current top tier of frontier models, but it is not the strongest performer across every task type.
How do you access Qwen3.8-Max today, and what does it cost?
Qwen3.8-Max is available now through Alibaba Cloud’s Model Studio API and through the QwenWork platform, which entered public beta on August 2, 2026. Qwen3.8-Max costs $2.00 per million input tokens and $6.00 per million output tokens, with implicit caching at $0.25 per million tokens. To put that in context, that $8 combined rate is less than a third of Claude Opus 5’s $30 and under a quarter of GPT-5.6 Sol Standard mode at $35. Prices are as of August 3, 2026, and should be confirmed on Alibaba’s official pricing page before building any cost model around them.
The API is OpenAI- and DashScope-compatible, so integration requires only a base-URL and model-ID change for teams already using OpenAI-format tooling. It also supports an Anthropic-compatible interface, so it works with tools like Claude Code, Codex, and Qwen Code. For non-developer users, Alibaba’s QwenWork workplace AI platform provides a browser-based interface for coding, document work, and collaborative tasks. See the recent OpenAI GPT-5.6 Luna price cut coverage for context on how frontier API pricing has been shifting across the major providers.
When will the open weights be available, and what hardware do you need?
Alibaba has committed to publishing the open weights of Qwen3.8-Max during the week of August 10, 2026, on Hugging Face and ModelScope. Alibaba has said it will publish weights for Qwen3.8-Max and a new Qwen3.8-27B during the week of August 10, 2026, on Hugging Face and ModelScope. That would be the first open-weight release at Max scale for the Qwen line.
As of August 8, that release has not yet occurred. At the time of writing, no matching repository appears on the Qwen organization page at Hugging Face. The smaller companion model, Qwen3.8-27B, is also going open-weights alongside the flagship. The 27B model is expected to be far more practical for self-hosting because the full 2.4-trillion-parameter flagship is a multi-node datacenter artifact requiring substantial GPU memory to serve. Qwen 3.5 and 3.6 shipped under Apache 2.0, and Qwen3.6-27B carries an Apache 2.0 LICENSE file on Hugging Face. That is a strong precedent, but not a commitment for Qwen3.8. Read the actual LICENSE file when it appears before building any commercial workflow on top of the open weights. For broader context on how open-weight AI models are reshaping developer options, see our coverage of changes inside Google’s Gemini ecosystem and the recent leadership shift at Google DeepMind.
What are the data-privacy and security risks for US users?
US users and businesses should treat the hosted Qwen3.8-Max API as a Chinese-jurisdiction service and keep sensitive data out of it. This is not a theoretical concern about a foreign company, it is a structural legal reality that applies to any platform operated under Chinese law.
QwenWork is an Alibaba product subject to Chinese law, including the Cybersecurity Law, the Data Security Law, and the Personal Information Protection Law. Under these frameworks, Alibaba can be compelled to provide access to data processed through its systems to Chinese government authorities. This is not a theoretical risk, it is a structural reality of operating a platform under Chinese jurisdiction.
For organizations handling proprietary trade secrets, regulated data (HIPAA, GDPR, financial records), sensitive customer information, or defense-related work, routing data through QwenWork carries compliance and security exposure that must be formally assessed before adoption. The practical guidance for US individuals and businesses is straightforward: general-purpose tasks such as summarizing publicly available documents, drafting non-confidential text, or exploring the model’s reasoning capabilities carry a lower risk than sending personal health information, legal documents, financial data, or internal business communications through the hosted API.
The main concern for US businesses is data jurisdiction when using the hosted API. If you self-host Qwen’s open weights, that concern goes away. Once the open weights are published and the license is confirmed, self-hosting on US-based infrastructure becomes an option that eliminates the jurisdiction question entirely. Stop and involve your legal and compliance team before integrating the hosted API into any workflow that handles regulated or proprietary data. For related context on data and AI privacy considerations, see our piece on ChatGPT connecting to bank accounts and what that privacy picture looks like.
How does Qwen3.8-Max fit into the broader AI competition?
Qwen3.8-Max arrives in the middle of a rapid push by Chinese AI labs into the multi-trillion-parameter tier, and it represents the most capable open-weight challenger to US frontier models released so far in 2026. Qwen3.8-Max sits close to the size of domestic rival Moonshot AI’s Kimi K3, which launched last month with 2.8 trillion parameters. The new Qwen3.8-Max ranks higher on several benchmarks than Kimi K3, though the comparison is contested depending on which benchmark is used.
The competitive significance for US consumers and developers is primarily about pricing leverage and open-weight availability. The open weights will be the first time Alibaba has open-sourced a model at this scale. That matters because it gives developers, researchers, and smaller organizations access to a frontier-class model they can run on their own infrastructure without per-token fees or data-jurisdiction concerns. The AI model landscape is also shifting at a governance level; our coverage of the EU’s new AI labeling rules provides relevant context on how regulations are beginning to follow this rapid scaling race. For a sense of the user adoption scale these models compete for, see our piece on ChatGPT approaching 1 billion weekly users.
Alibaba stock rose 4.5% in premarket trading in New York and 7% on the Hong Kong exchange following the announcement, which signals that investors view the launch as a meaningful competitive development. That said, stock movement reflects market sentiment, not a validation of the model’s technical claims, and independent benchmark results will be the more useful signal once they accumulate over the coming weeks.
FAQ
Is Qwen3.8-Max safe to use for everyday tasks?
Qwen3.8-Max is generally safe for general-purpose tasks such as drafting text, summarizing public documents, writing code, or analyzing images you do not mind being processed by an Alibaba-operated service. Keep personal, financial, health, and work-sensitive data out of the hosted API because prompts are processed under Chinese data jurisdiction, where Chinese government authorities can compel access under the Cybersecurity Law and the Data Security Law. The safety risk is about data exposure, not the model’s conversational behavior.
What is the difference between Qwen3.8-Max and the open weights?
Qwen3.8-Max is currently available only as a hosted API through Alibaba Cloud’s QwenCloud at $2 per million input tokens and $6 per million output tokens. The open weights, the actual model files that let anyone run the model on their own hardware, are scheduled for release on Hugging Face and ModelScope during the week of August 10, 2026, but had not appeared as of August 8. Self-hosting the open weights removes the data-jurisdiction concern entirely, though running the full 2.4-trillion-parameter model requires multi-node datacenter infrastructure.
How does Qwen3.8-Max compare to ChatGPT and Claude?
Qwen3.8-Max is competitive with GPT-5.6 Sol Max and Claude Fable 5 on several benchmarks, leading on OSWorld-Verified (agentic computer use) and PaperBench (research reproduction) according to Alibaba’s self-reported scores. Claude Fable 5 leads on software engineering benchmarks including SWE-Pro, and GPT-5.6 Sol leads on others including TerminalBench-2.1, per the same Alibaba-published comparison. Independent, third-party verification of the full benchmark picture is still limited as of August 8, 2026, so treat the current rankings as a snapshot rather than a settled verdict.
Can developers integrate Qwen3.8-Max into existing apps?
Yes, and the integration is straightforward for teams already using OpenAI-compatible or Anthropic-compatible tooling because Qwen3.8-Max supports both API formats with only a base-URL and model-ID change required. The model is available under the model ID qwen3.8-max on QwenCloud and is also accessible through Vercel’s AI Gateway. Developers building compliance-sensitive applications should complete a data-jurisdiction review before routing production data through the hosted endpoint.
What is the Qwen3.8-27B and who is it for?
Qwen3.8-27B is a smaller, 27-billion-parameter companion model that Alibaba plans to release as open weights alongside Qwen3.8-Max during the week of August 10, 2026. It is aimed at developers and researchers who want to run a capable Qwen3.8-generation model on a single high-end GPU or a small local cluster rather than multi-node datacenter hardware. The license for the 27B model has not been formally disclosed as of August 8, so confirm the LICENSE file on Hugging Face before building any commercial workflow on top of it.