Qwen3.8-Max arrives on Alibaba Cloud, offering massive context capacity for enterprise workloads

Alibaba's Qwen3.8-Max uses 2.4 trillion total parameters, activates 95 billion during inference, and supports a one-million-token context window.

Alibaba has released Qwen3.8-Max, a new multimodal AI model with 2.4 trillion total parameters and support for context windows of up to one million tokens.

The model is available through APIs on Alibaba Cloud Model Studio, and Alibaba said the model weights are scheduled for release the week of August 10. Qwen3.8-Max is also available through QwenWork, the company's workplace AI agent platform.

Model Studio provides Qwen models through Alibaba's native APIs and OpenAI-compatible APIs. Alibaba Cloud documentation states that applications using OpenAI client libraries can connect by changing the API key, base URL, and model name.

Alibaba said Qwen3.8-Max currently ranks fifth on Text Arena, second on Vision Arena, and fourth on Frontend Code Arena. Frontend Code Arena evaluates models on frontend development tasks.

Arena rankings are based on pairwise user preferences rather than a fixed set of benchmark questions, with users comparing responses from two models and voting for the stronger result.

Sparse architecture and longer context

Qwen3.8-Max builds on the architecture used in Qwen 3.5 and uses a sparse Mixture-of-Experts model alongside a hybrid attention mechanism.

Although the model contains 2.4 trillion parameters in total, only 95 billion are activated during inference. Mixture-of-Experts architectures route inputs through selected parts of a model, reducing the amount of computation required for each token compared with activating the full parameter set.

Alibaba has not published comparable Qwen3.8-Max figures for inference latency, throughput, or hardware requirements in the materials reviewed for this article.

The model also supports multimodal inputs, allowing it to process text and visual information within the same system.

Qwen3.8-Max's one-million-token context window places it alongside existing Qwen models that support inputs of that size. Alibaba Cloud currently lists one-million-token support for Qwen3.7-Max and some Qwen Plus variants on Model Studio.

Context length can also affect deployment costs. Alibaba Cloud uses tiered pricing for some Qwen models, with per-token rates varying according to the number of input tokens in a request, although pricing for Qwen3.8-Max has not yet been listed in the documentation reviewed for this article.

The size of a context window indicates how much information a model can accept in a single request, but not how consistently it can use information across the entire input. Research on long-context models has found that performance can vary depending on where relevant information appears within the context window.

Alibaba said Qwen3.8-Max can process material including documents running to hundreds of pages, television programs, and lengthy livestream recordings. The company said users can then search and query information from those inputs through the model.

Coding and multimodal tasks

Alibaba has tested Qwen3.8-Max on software development tasks that require the model to operate across extended periods.

In one internal test, the company said the model worked on a software engineering project for 16 days without human intervention. The task involved developing a self-evolving agent framework that incorporated user feedback, community practices, and results from its own testing.

The model repeatedly generated code, ran tests, produced previews, and analyzed logs during the project. The resulting framework, called "oh-my-cli," has been released as open-source software on GitHub.

Alibaba also said Qwen3.8-Max was tested on research tasks requiring it to reproduce experiments described in published papers. In some cases, according to the company, methods generated by the model achieved better results than those reported in the original research.

The model also participated in the WWW2025 Multimodal Dialogue Intent Recognition Challenge. Alibaba said Qwen3.8-Max recorded a higher result than human participants in the challenge, which evaluates systems on their ability to identify user intent from multimodal customer-service interactions.

Alibaba has also tested Qwen3.8-Max on workloads that involve multiple stages and constraints, including legal document review, financial research, and architectural modeling.

The company said it used reinforcement learning environments to train the model on agent-based tasks involving planning, execution, evaluation, and subsequent adjustments.

Its multimodal capabilities also allow it to perform tasks based on visual feedback. Alibaba demonstrated the model rebuilding frontend projects from screenshots, converting 2D floor plans into 3D interior models, editing video footage, and generating interactive applications from natural-language instructions.

The company introduced a benchmark called RecreationBench to evaluate this type of visual application development.

In the benchmark, the model operates in a black-box environment without internet access or access to an application's source code. It interacts with a running application, observes its behavior and visual output, and attempts to recreate it from scratch.

Alibaba said Qwen3.8-Max used repeated interaction and visual feedback to modify its implementation during these tests, combining software development with visual evaluation across multiple steps.