Preface¶

On August 3, 2026, Alibaba Cloud officially released Qwen 3.8-Max, the most powerful model in the Qwen series to date. Unlike previous Max-level models that only offered API access, Alibaba announced that the model weights would be released on Hugging Face and ModelScope next week – this marks the first time a Qwen-Max level model has been open-sourced.

For developers, several highlights of this release are worth noting: 2.4 trillion total parameters, a sparse MoE architecture with 95B activated parameters, a 1 million token context window, and upgraded capabilities for Agent programming and long-range tasks. The model is now live on QwenCloud and Vercel AI Gateway, with APIs available for immediate use; the open-sourced weights are expected to launch in one week.

This article is based on Alibaba Cloud’s official blog and developer community reports, sorting out verified technical specifications, capability boundaries, and access methods.

Core Specifications: 2.4T MoE and 1M Context Window¶

Qwen 3.8-Max is evolved based on the Qwen 3.5 architecture, adopting Sparse Mixture of Experts (Sparse MoE) and hybrid attention mechanisms. The key metrics disclosed by the official are as follows:
- Total parameters: 2.4 trillion (2.4T)
- Activated parameters: Approximately 95B per inference
- Context window: Up to 1 million tokens
- Modalities: Text + Vision (multimodal input)
- Architecture: Sparse MoE, balancing scale and inference efficiency

In public evaluations, the official stated that Qwen 3.8-Max ranks 5th in Text Arena, 2nd in Vision Arena, and 4th in Frontend Code Arena. These rankings come from Alibaba’s official press release, and the specific details of the leaderboards shall prevail on each Arena platform.

The significance of the MoE design lies in that the total number of parameters can be very large, but only a portion of experts are activated each time, making the inference cost relatively controllable. For Agent scenarios requiring long context and complex reasoning, the 1M token window means that entire codebases, long documents, and multi-turn conversations can be processed in a single request without frequent truncation or segmentation.

First Open-Sourcing of Max-Level Weights¶

Previously, the open-source Qwen series mainly included Qwen2, Qwen2.5, Qwen3, etc. The Max-level flagship models had long only been available via API. Qwen 3.8-Max is the first Max-level model that the official has clearly stated will release its weights.

Current Status (as of 2026-08-04):

For teams performing self-hosting, fine-tuning, or private deployment, the open-sourcing of weights is the headline of this release. A 2.4T-scale MoE model has extremely high computing power and storage requirements (the community estimates the download size may be in the terabyte range), and cluster resources need to be evaluated before deployment; but the fact that “Max-level models can run locally” will change the technology selection of many teams.

Agent Programming and Long-Range Tasks¶

The official blog uses multiple long-range cases to illustrate the model’s positioning. The core is not “writing a single function”, but completing multi-day, multi-step end-to-end tasks without human intervention.

Autonomous Programming: oh-my-cli Project¶

Qwen 3.8-Max was asked to create the oh-my-cli project from scratch and build a self-evolving harness during a long-term run of more than 10 days. According to official disclosure, as of July 30, 2026, after about 16 days of fully automated operation, the repository has accumulated 265 commits, 127 PRs, and 151 issues. The full trace is publicly available on GitHub: qwen-code-dev-bot/oh-my-cli.

Paper Replication and Improvement¶

After receiving the paper Unified Data Selection for LLM Reasoning, the model autonomously replicated the experiment in about 5 days (~125 hours) without starter code, wrote approximately 7600 lines of code, and outperformed the original paper’s method on mathematical benchmarks such as AIME24.

Multimodal Competition¶

In the WWW2025 Multimodal Dialogue Intent Recognition Challenge (with 526 human teams participating), the model independently completed the solution within the 24-hour limit, with a final accuracy of 0.853, exceeding 87% of the human teams. These cases come from official tests, and the actual effect varies by task; developers are advised to conduct their own evaluations.

For daily development, a more direct starting point is that the official has provided docking configurations with mainstream Agent tools, including Claude Code, Codex, Qwen Code, OpenClaw, Qoder CLI, etc.

How to Access: API and Gateway¶

1. QwenCloud (OpenAI Compatible)¶

After obtaining your DASHSCOPE_API_KEY at home.qwencloud.com, you can use the OpenAI SDK to make calls. The official example:

from openai import OpenAI
import os

client = OpenAI(
    api_key=os.environ.get("DASHSCOPE_API_KEY"),
    base_url=os.environ.get(
        "DASHSCOPE_BASE_URL",
        "https://dashscope-intl.aliyuncs.com/compatible-mode/v1",
    ),
)

completion = client.chat.completions.create(
    model="qwen3.8-max",
    messages=[{"role": "user", "content": "Write a Python function to merge two sorted linked lists."}],
    extra_body={"enable_thinking": True},
    reasoning_effort="xhigh",  # Optional: xhigh (default), medium, low
    stream=True,
)

reasoning_effort is used to adjust the inference depth and cost: xhigh is suitable for complex tasks, while low prioritizes speed and cost. The official states that preserve_thinking is enabled by default.

2. Claude Code¶

If you already use Claude Code, you can switch models through the Anthropic-compatible endpoint without modifying your workflow:

npm install -g @anthropic-ai/claude-code

export ANTHROPIC_MODEL="qwen3.8-max"
export ANTHROPIC_SMALL_FAST_MODEL="qwen3.8-max"
export ANTHROPIC_BASE_URL=https://dashscope-intl.aliyuncs.com/apps/anthropic
export ANTHROPIC_AUTH_TOKEN=<your_api_key>

claude

3. Vercel AI Gateway¶

Vercel AI Gateway already supports this model. If your project uses the Vercel AI SDK, simply change the model parameter to alibaba/qwen3.8-max; the Gateway will handle routing, usage statistics, and failover. For scenarios already running coding agents on Vercel or requiring multimodal input, the migration cost is low.

4. Waiting for ModelScope / Hugging Face Weights¶

Teams planning to self-host should follow the official Qwen accounts on ModelScope and Hugging Face. The weights have not been released yet, so do not hardcode deployment scripts in advance; after the release, prepare the MoE inference framework and hardware according to the model card instructions.

Pricing and Selection Reference¶

According to the QwenCloud official pricing page and third-party summaries, the API pricing is approximately:
- Input: \(2 / million tokens** - Output: **\)6 / million tokens

The 1M context window is billed at a unified unit price, so pay attention to token usage for long-context tasks. Compared to closed-source APIs, once the Max-level open-source weights are launched, self-hosting will provide new options for scenarios such as data compliance and batch inference; however, the deployment threshold for a 2.4T MoE model is not low, and most teams will still mainly use APIs or Gateways in the short term.

Summary¶

Qwen 3.8-Max pushes the flagship capabilities of Qwen to 2.4T MoE + 1M context window, and for the first time promises to open-source Max-level weights. For developers:
1. Available now: APIs are provided on QwenCloud, Model Studio, and Vercel AI Gateway.
2. Agent toolchain is ready: Official configuration examples are available for Claude Code, Codex, Qwen Code, etc.
3. Watch for weights next week: After the open-sourcing on Hugging Face / ModelScope, self-hosting and fine-tuning will truly be unlocked.

It is recommended to first use the API to run tests on your own benchmarks (coding, long documents, multimodal Agents) before deciding whether to wait for the weights for private deployment. Official blog posts and API documentation will be updated continuously, please refer to the Alibaba Cloud Qwen 3.8-Max release article as the authoritative source.