Preface¶

On August 3, 2026, Alibaba officially launched the Qwen3.8-Max, the most powerful model in the Qwen series to date, in Hangzhou. This is a Sparse Mixture of Experts (Sparse MoE) model with a total parameter count of 2.4 trillion and approximately 95 billion activated parameters per inference run. It supports a maximum context window of 1 million tokens, and has gained over 1,090 points and nearly 600 comments on Hacker News, making it one of the most talked-about open-source AI topics in the developer community recently.

Unlike the previous Qwen series’ rhythm of “API first, weights later”, Alibaba announced that it will publicly release the Max-level model weights on Hugging Face and ModelScope for the first time during the second week of August (around August 10). It will also open-source the smaller Qwen3.8-27B checkpoint. For developers focused on long-range autonomous programming Agents, this release deserves careful breakdown.

Model Overview: 2.4T MoE and Million-Token Context¶

Qwen3.8-Max is evolved based on the Qwen 3.5 architecture, adopting a Sparse MoE + Hybrid Attention design: with a total parameter size of 2.4T, only about 95B parameters are activated per request, balancing scale and inference cost.

The core specifications disclosed by the official are as follows:

On public leaderboards, Qwen3.8-Max ranks 5th in Text Arena, 2nd in Vision Arena, and 4th in Frontend Code Arena. Alibaba also launched the office Agent platform QwenWork at the same time, where users can directly experience the model’s capabilities.

16-Day Autonomous Programming: The oh-my-cli Case¶

What attracted the most attention from developers this release was not just the benchmark scores, but the internal test of long-horizon autonomous software engineering disclosed by Alibaba.

According to official blog posts and multiple media reports, Qwen3.8-Max built a self-evolving Agent framework named oh-my-cli from scratch in about 16 days with no human intervention. The model independently established an engineering closed loop of “user feedback → code generation → testing → preview → log analysis” and continued iterating.

The GitHub repository of this project has been made public: https://github.com/qwen-code-dev-bot/oh-my-cli

As seen in the README, oh-my-cli is a code Agent CLI based on Node.js 22 + TypeScript, with core features including:
- Compatible with any OpenAI-compatible endpoints (including DashScope)
- Default fail-closed security approval and command policies
- JSONL persistent sessions supporting resume / compact / undo operations
- Headless JSON event stream for easy automated integration
- Comes with Electron Desktop and Web Delivery Board

The repository uses the Apache 2.0 license, and includes an AUTONOMY.md autonomy contract document, proving that its “self-evolution” is not a marketing slogan, but an engineering practice with clear governance boundaries.

It is important to emphasize: the 16-day autonomous programming result is Alibaba’s internal test conclusion, which has not been independently reproduced by third parties; it should be regarded as a demonstration of capability direction rather than a verified general productivity benchmark.

API Access: OpenAI Compatibility and Claude Code Integration¶

Qwen3.8-Max is now open to global developers via the Alibaba Cloud Model Studio (QwenCloud) API. The official supports two protocols: OpenAI Chat Completions and Anthropic API, which means minimal changes to existing toolchains – usually only the base_url and model need to be replaced to access the service.

According to media reports such as the-decoder, the following tools can be directly connected:
- Claude Code
- Codex / Qoder CLI
- Qwen Code
- OpenClaw

The model also provides the reasoning_effort parameter, which allows switching between three levels of reasoning intensity between speed and sub-task thoroughness.

Quick Access Example¶

Calling via the OpenAI-compatible method (replace with your API Key and endpoint):

from openai import OpenAI

client = OpenAI(
    api_key="YOUR_DASHSCOPE_API_KEY",
    base_url="https://dashscope.aliyuncs.com/compatible-mode/v1",
)

response = client.chat.completions.create(
    model="qwen3.8-max",
    messages=[
        {"role": "user", "content": "Review this Python module for edge cases."}
    ],
    extra_body={"reasoning_effort": "medium"},
)

print(response.choices[0].message.content)

If using Claude Code, you only need to specify the Anthropic-compatible endpoint and model ID in the configuration to run Qwen3.8-Max as the backend model without rewriting Agent logic.

Pricing and Rate Limits (Official Model Page)¶

Citing data from Alibaba’s model page via MarkTechPost, the pricing for Qwen3.8-Max API is approximately:
- Input: \(2 / million tokens** - Output: **\)6 / million tokens
- Cached Input: $0.25 / million tokens

Rate limits: 2 million tokens per minute, 15,000 requests per minute. Compared to similar closed-source flagship models, the price is relatively competitive, but the actual cost needs to be measured based on task length and reasoning mode.

Open-Source Weights: First-Ever Max-Level Release¶

Alibaba clearly stated that Qwen3.8-Max weights will be launched on Hugging Face and ModelScope in the week following the release (around August 10). This is the first time the Qwen family has open-sourced weights for a Max-level model.

The Qwen3.8-27B will also be open-sourced simultaneously – a checkpoint more suitable for local deployment on ordinary GPU clusters. It is important to soberly recognize that:
- The 2.4T full weights are a multi-node data center-level artifact, which almost cannot be fully loaded on a personal workstation
- As of the time of writing, details such as the open-source license filename and exact activation parameter values have not been fully disclosed
- For most teams, API calls + 27B local deployment is a more realistic combination

Community Response: Calm and Expectation on HN¶

The Hacker News post “Qwen3.8-Max: A New Bar for Coding and Cowork” (ID: 49150470) has gained 1,090 points and 594 comments, with discussions focusing on several directions:
1. The credibility and reproducibility of the 16-day autonomous programming – some comments requested the release of full logs and intermediate artifacts
2. The significance of 2.4T MoE open-sourcing for the self-hosted ecosystem – some people expect distillation and community fine-tuning similar to DeepSeek
3. Claude Code compatibility reducing migration costs – some developers have shared their initial experience after replacing the endpoint
4. The gap between Chinese and US models – some discussions compared Qwen3.8-Max horizontally with flagships such as Opus and GPT-5.x

Overall, the community’s attitude is “the capability statement is impressive, but we will wait for the weights and third-party evaluations before drawing conclusions”.

Practical Recommendations for Developers¶

If you are evaluating Qwen3.8-Max, you can proceed along the following path:
1. Start with the API: Register on Model Studio, use the OpenAI SDK or Claude Code to verify small-scale coding / Agent tasks
2. Follow oh-my-cli: Read its AUTONOMY.md and session JSONL design to understand the engineering closed loop of long-range Agents
3. Try the 27B model first after weights are released: Local deployability is far better than the 2.4T full checkpoint
4. Test long-context scenarios: The million-token window has potential value for codebase-level RAG and long-document analysis, but pay attention to latency and costs
5. Stay prudent: Internal benchmarks and the 16-day case are directional signals; production environments still require stress testing with your own datasets

Summary¶

The release of Qwen3.8-Max marks Alibaba’s simultaneous efforts in three areas: ultra-large-scale MoE, million-token context, and long-range autonomous Agents. The 16-day human-free construction of oh-my-cli case, the promise of first-ever open-sourcing of Max-level weights, and the dual-protocol compatibility of Claude Code / OpenAI have collectively raised developer expectations.

In the next one to two weeks, the weight launch, license details, and independent evaluations will be unveiled one after another. For readers who focus on coding Agents and open-source large models, this is undoubtedly one of the most值得跟踪的 technical events in August 2026.