Alibaba’s Qwen3.8-Max brings 2.4 trillion parameters, massive context, multimodal AI and open weights to the global AI landscape.
Key Highlights
- 2.4 Trillion Parameters
- 1M Token Context
- Open Weights Released
- Multimodal AI Support
- Advanced Agent Features
- Custom Commercial License
Alibaba’s Qwen team has made a major move in the AI race with Qwen3.8-Max, its latest flagship model. The model launched officially on August 3, 2026, and its open weights arrived days later.
With 2.4 trillion total parameters, a 1-million-token context capability, multimodal input, and stronger agentic skills, Qwen3.8-Max is built for complex work rather than simple chatbot conversations.
Qwen3.8-Max Gets a Massive MoE Architecture
The biggest headline is the model's size. Qwen3.8-Max contains about 2.4 trillion parameters and uses a Mixture-of-Experts architecture.
However, it does not activate all parameters for every token. Around 95 billion parameters are active per token, helping the model handle its huge overall capacity more efficiently. The open-weight model is identified as Qwen3.8-2.4T-A95B.
This makes Qwen3.8-Max one of the largest publicly released AI models available today.
One Million Tokens for Long Tasks
Another major feature of Qwen3.8-Max is its extremely large context capability.
The model natively supports a 262,144-token context, while the context can be extended to about 1.01 million tokens.
That gives Qwen3.8-Max room to work with very large documents, software repositories, research material, and long-running projects without constantly losing earlier information.
For developers and researchers, this can be especially useful for large-scale coding and document-based workflows.
Multimodal AI Is Built In
Qwen3.8-Max is not restricted to text. It also supports visual input.
The model can work with images and video alongside text. This expands its use beyond normal chat and makes it suitable for visual analysis, research, document processing, and agent workflows.
Qwen is also positioning the model around coding, professional work, scientific research, and long-horizon agent tasks.
Qwen3.8-Max Focuses on AI Agents
The bigger goal behind Qwen3.8-Max is autonomous work.
Qwen says the model was designed to handle tasks that can continue for hours, days, or even weeks with less human intervention. In one reported test, the model carried out autonomous programming for about 16 days. It also worked on research reproduction and a virtual e-commerce business across more than 2,000 interactions.
These examples show why Qwen3.8-Max is being positioned as an agent-focused model, not just another general-purpose chatbot.
Benchmark Results Show Strong Performance
The published results for Qwen3.8-Max are impressive in several areas.
The model scored 93.0 on PaperBench, 86.1 on OSWorld-Verified, and 91.5 on a parametric CAD benchmark, according to results reported from Alibaba's benchmark table. However, it did not lead every test. Its 67.7 SWE-bench Pro score was below some competing models, while its 68.3 VideoMME v2 score also trailed the reported leader.
These figures should be viewed carefully because the published benchmark results are primarily from Qwen and independent evaluations are still developing.
Open Weights Change the Qwen Strategy
The biggest recent development came on August 13-14, when Alibaba released the weights for Qwen3.8-2.4T-A95B. This marked the first time a Qwen Max-class flagship model received an open-weight release.
That gives developers more freedom to experiment with Qwen3.8-Max, although running the full model locally remains extremely demanding.
Reports say the original model weighs around 4.9TB, while an Unsloth quantization approach reduced it to about 397GB. That still requires substantial hardware for local use.
The License Needs Attention
There is an important catch with the open release.
The flagship Qwen3.8-Max weights do not use the Apache 2.0 license. They are distributed under the separate Qwen3.8-Max License. Companies running certain large AI services can face additional licensing requirements, including a separate license when qualifying Model-as-a-Service or AI Work Assistant businesses exceed $50 million in aggregate revenue over a consecutive 12-month period.
Meanwhile, the smaller Qwen3.8-27B model uses Apache 2.0. Developers should therefore check the exact license before using Qwen3.8-Max in a commercial product.
Qwen3.8-Max Pricing and Availability
For API users, reported international pricing is $2 per million input tokens and $6 per million output tokens, with reported implicit cache pricing of $0.25 per million tokens. Mainland China pricing is reported at 12 yuan for input and 36 yuan for output per million tokens.
The model is also available through Qwen's ecosystem. Qwen Studio currently lists Qwen3.8-Max, giving users a direct way to access the model online.
Also Read: Qwen AI - How To Get Started? Complete Guide For Beginners
Qwen3.8-Max Could Be a Major AI Release
Qwen3.8-Max is significant because it combines enormous model capacity with long-context processing, multimodal input, reasoning, coding, and autonomous agent capabilities.
Its open-weight release makes the announcement even more important. Still, the custom license and demanding hardware requirements mean it is not a simple drop-in replacement for smaller open models.
For users, developers, and businesses worldwide, Qwen3.8-Max is now one of the most closely watched AI models of 2026. Its next challenge will be proving that its strong published results translate into reliable performance across independent real-world testing.
Also Read: Is Qwen AI Free? You Should Know Before Use It