Qwen3.8-Max Lands as an Open-Source 2.4T MoE Model, and a Local AI Gateway Makes It a Drop-In Replacement for Claude Code
On August 3, 2026, Alibaba's Tongyi Qianwen team officially released Qwen3.8-Max. This is the largest and most capable model in the Qianwen family to date, and it marks the first time Qianwen has open-sourced a Max-level flagship model. The open-source weights are expected to be available for public download via Hugging Face and ModelScope during the week of August 10.
The release of Qwen3.8-Max signifies two things. First, the model's capabilities have genuinely reached the global frontier. Second, and more relevant to us as developers, as frontier models become more numerous, more powerful, and update faster, how should our toolchains and workflows adapt?
This article will start from the technical specifications of Qwen3.8-Max, combine the pain points of multi-model access in actual development, and introduce how to complete the access and management of new models without modifying any application code through a full-featured AI Gateway.
Qwen3.8-Max Technical Specifications: The First Open-Source Max-Level Flagship
Qwen3.8-Max uses a Sparse Mixture of Experts (Sparse MoE) architecture, with a total parameter count of 2.4 trillion and 95 billion parameters activated per inference. The model is significantly scaled up based on the Qwen 3.5 architecture, supports a maximum context window of 1M tokens, and possesses visual understanding capabilities.
It is necessary to emphasize a landmark event in the history of Chinese large model development: Qwen3.8-Max is the first time Qianwen has open-sourced a Max-level model. Before this, Qianwen's open-source releases were all smaller-spec versions, with the Max-level flagship only available via API. This open-sourcing breaks that convention; Alibaba chose to make the flagship model's weights public to the community, which is a significant push for the entire open-source AI ecosystem.
Performance and Pricing
In various benchmark tests, Qwen3.8-Max demonstrated comprehensive improvement over its predecessor, Qwen3.7-Max, and approached or even surpassed top-tier overseas closed-source models in multiple dimensions.
Below is a comparison of some representative Benchmark data:
| Benchmark | Claude Opus 4.8 | Claude Fable 5 | GPT 5.6 Sol | Qwen3.8-Max |
|---|---|---|---|---|
| PaperBench | 80.3 | 88.8 | 90.5 | 93.0 |
| FrontierSWE | 70.0 | 88.8 | — | 73.5 |
| Terminal Bench 2.1 | 84.6 | 84.6 | 88.8 | 86.6 |
| IFBench | 62.2 | 63.5 | 72.7 | 82.8 |
| CoWorkBench | 72.3 | 75.9 | 71.5 | 74.8 |
| GPQA Diamond | 92.0 | 92.6 | 94.1 | 92.6 |
On Chatbot Arena's latest Frontend Code leaderboard, Qwen3.8-Max is only 1 point behind Claude Opus 5 High. In the Text Arena, Qianwen also closely follows Anthropic, ranking second globally.
In terms of pricing, the API price for Qwen3.8-Max is 12 RMB per million input tokens and 36 RMB per million output tokens domestically, with cache hit input costing only 1.5 RMB. Internationally, the input and output prices are only 40% and 24% of Claude Opus 5's, respectively. Combined with the "Temu Model" wave sparked by DeepSeek V4 Flash, Chinese models have formed a clear advantage in the cost-performance dimension.
Impressive Coding Capability
The most impressive case from Qwen3.8-Max's release was a 16-day fully autonomous coding test. The model started from an empty folder and independently completed the construction of the oh-my-cli project, accumulating 265 commits, 127 PRs, and 151 Issues, and autonomously built a self-evolving Harness framework.
In another test, Qwen3.8-Max independently reproduced the complete experimental workflow of an academic paper (Unified Data Selection for LLM Reasoning) in about 5 days without any initial code, writing approximately 7,600 lines of code and completing 33 rounds of GPU training. Ultimately, it not only reproduced the paper's core findings but also autonomously proposed and validated 18 improvement plans, surpassing the original paper's method by 2.7 percentage points on the AIME24 competition-level math benchmark.
These results indicate that Qwen3.8-Max's performance in long-duration autonomous programming tasks is already quite competitive, and it is more than sufficient for daily software development work.
The Old Problem of Accessing New Models
Performance and price are satisfactory, but returning to actual development scenarios, accessing a new model is not just about getting an API Key.
Take the two most mainstream coding assistants as an example. Claude Code uses the Anthropic protocol, and Codex uses the OpenAI Responses protocol. Qwen3.8-Max's API is compatible with the OpenAI Chat Completions protocol, and Alibaba also provides an Anthropic-compatible endpoint.
If a developer wants to use Qwen3.8-Max directly in Claude Code, they need to modify a set of environment variables according to the official documentation:
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=sk-your-dashscope-key
claude
This method works, but it brings several practical problems:
First, the configuration is hardcoded. Once this set of environment variables is set, Claude Code can only call Qwen3.8-Max. To switch back to Claude or try DeepSeek, you have to change the environment variables and restart the session.
Second, Keys are exposed. Each project and each tool directly holds the real API Key. As projects multiply, Keys are scattered across various
.envfiles, shell profiles, and configuration files, increasing management costs and leakage risks simultaneously.Third, there is no global perspective. When using multiple models and multiple projects simultaneously, questions like how much was spent on AI this month, which project consumed the most, and which model offers better cost-performance are almost impossible to answer without a unified management layer.
In 2024, most developers' .env files might have only one line: OPENAI_API_KEY. By August 2026, times have changed:
OPENAI_API_KEY=sk-xxxx
ANTHROPIC_API_KEY=sk-ant-xxxx
DASHSCOPE_API_KEY=sk-dash-xxxx
DEEPSEEK_API_KEY=sk-deep-xxxx
GOOGLE_API_KEY=AIza-xxxx
Every time a new model is released, this list grows by one line. Each additional line adds another layer of management complexity.
Protocol Conversion and Model Mapping in AI Gateway
The solution to this type of problem is already very mature in traditional web architecture: add a gateway layer between the client and backend services. In microservice architecture, an API Gateway handles routing, authentication, rate limiting, and monitoring. An AI Gateway does the same thing, except the backend services are replaced by the APIs of various AI models.
Two key capabilities of an AI Gateway directly address the pain points mentioned earlier.
Protocol Conversion
Currently, there are three mainstream protocols in the AI API space: OpenAI's Chat Completions, Anthropic's Messages, and Google's Gemini API. The three differ in request format, response structure, streaming output, and tool calling.
Protocol conversion means that after receiving a client request, the gateway automatically converts it to the protocol format required by the target model and converts the response back to the format expected by the client. This way, regardless of which protocol the upstream coding assistant uses, and whether the downstream model is Qwen (OpenAI protocol) or Claude (Anthropic protocol), the application layer does not need to care or make any adaptations.
Model Mapping
Coding assistants specify a model name when making a request, for example, Claude Code requests claude-opus-5. Model mapping allows this name to be redirected to another model at the gateway layer. For instance, mapping claude-opus-5 to qwen3.8-max means requests sent by Claude Code are transparently forwarded by the gateway to Qwen3.8-Max's API, with the entire process being completely transparent to the client.
Combining these two capabilities achieves one effect: the coding assistant side doesn't need to change anything; the gateway layer handles all protocol adaptation and model routing.
Practical Access to Qwen3.8-Max via ServBay AI Gateway
ServBay AI Gateway is a full-featured AI gateway built into ServBay, running on the developer's local machine. It supports adding various official AI APIs, subscription accounts, and various third-party relay stations as upstream channels.
The entire process of accessing Qwen3.8-Max with ServBay AI Gateway can be broken down into three steps.
Step 1: Add Qwen3.8-Max as a New Channel
Add a new channel in ServBay AI Gateway and fill in the API information for Alibaba Cloud DashScope:
Channel type select OpenAI Compatible
Base URL fill in
https://dashscope.aliyuncs.com/compatible-mode/v1(domestic) orhttps://dashscope-intl.aliyuncs.com/compatible-mode/v1(international)API Key fill in the DashScope key
Add
qwen3.8-maxto available models
If targeting international users or needing Anthropic protocol compatibility, you can also separately add the Anthropic-compatible endpoint https://dashscope-intl.aliyuncs.com/apps/anthropic provided by Alibaba Cloud as another channel.
Step 2: Configure Model Mapping
In the gateway's model mapping rules, map claude-opus-5 to qwen3.8-max. This way, when Claude Code requests claude-opus-5, the gateway automatically routes the request to the Qwen3.8-Max channel and completes the protocol conversion.
Mapping is flexible and can be adjusted at any time. If you just want to try Qwen3.8-Max for a while, switching back only requires modifying the mapping rules in the gateway; the application layer is completely unaware.
Step 3: Point the Coding Assistant to the Gateway
Claude Code only needs to point to the local unified endpoint provided by ServBay AI Gateway, using the virtual Key assigned by the Gateway. Once this configuration is set, it no longer needs to be changed, regardless of how many models are connected to the backend or how routing strategies are adjusted.
Throughout the entire process, Claude Code does not know which model it is calling, nor does it need to know. It sends requests according to the Anthropic protocol, the gateway automatically completes the protocol conversion and sends the request to Qwen3.8-Max's OpenAI-compatible interface, then converts the response back to Anthropic format and returns it. The application layer's code and configuration are not modified at all.
This process is equally applicable to other coding assistants like Codex, Qoder, Qwen Code, and OpenClaw. Each assistant just needs to point to the local endpoint of the corresponding protocol on the Gateway, and the selection and switching of backend models are all completed at the gateway layer.
Multi-Model Coexistence: Channel Priority and Automatic Fallback
In actual development, developers rarely rely on a single model. A more common practice is to maintain access to multiple models simultaneously and allocate them flexibly based on task type and cost considerations.
ServBay AI Gateway supports setting priorities for different channels. A typical configuration is:
- High Priority: Qwen3.8-Max (low price, fast direct domestic connection)
- Medium Priority: DeepSeek V4 (extreme cost-performance, suitable for simple tasks)
- Low Priority: Claude Opus 5 (capability ceiling, fallback for complex tasks)
The gateway tries channels in order of priority. When a high-priority channel encounters a timeout, rate limit, or service anomaly, it automatically falls back to the next available channel, with the entire switching process transparent to the coding assistant. Developers can later check the statistics panel to see how many requests each channel actually handled and how many Tokens and costs each consumed.
Channel hot-switching is also a practical capability. Without needing to restart any services or interrupt ongoing coding sessions, you can directly enable or disable a channel or adjust the priority order in the Gateway management interface, and changes take effect immediately. When a model provider temporarily adjusts pricing or experiences service fluctuations, developers can complete strategy adjustments within seconds.
Virtual Keys and Usage Statistics
API Key management is an easily overlooked but impactful issue in multi-model, multi-project scenarios.
ServBay AI Gateway provides a virtual Key mechanism. Developers can create multiple virtual Keys and assign them to different projects or different team members. The real API Key is only stored inside the Gateway, encrypted, and never exposed to any downstream application.
The direct benefits of virtual Keys:
Each project uses an independent virtual Key, so usage statistics are naturally isolated by project.
If a virtual Key is accidentally leaked, simply revoke that Key; other projects are unaffected, and the real Key does not need to be changed.
At the end of the month, through the statistics panel, you can clearly see the request volume and cost for each virtual Key, each channel, and each model.
Combined with Qwen3.8-Max's low pricing advantage, this statistical capability can also answer a very practical question: if some traffic is migrated from Claude to Qwen3.8-Max, how much cost can actually be saved? The Gateway's statistics panel can provide an answer precise to the channel dimension.
More Possibilities After Qwen3.8-Max Open-Source Weights Release
The open-source weights for Qwen3.8-Max are expected to be released during the week of August 10. The complete 2.4 trillion parameter model has extremely high hardware requirements, and full local deployment is not realistic for individual developers. However, it is foreseeable that the community will soon release quantized and distilled versions to lower the hardware threshold for inference.
At that time, developers can simultaneously access the cloud-based Qwen3.8-Max API and a locally deployed quantized version in the Gateway, achieving load distribution through priority strategies. Tasks insensitive to latency and privacy go to the cloud, while tasks requiring fast response and data security go local. This hybrid cloud-local architecture is precisely the scenario where an AI Gateway can maximize its value.
Conclusion
The release of Qwen3.8-Max is a major event in the AI field for the second half of 2026. The first open-source Max-level flagship, performance benchmarking against top global closed-source models, and pricing only a fraction of competitors. For domestic developers, this is a very attractive new option.
But the model is just one link in the toolchain. Models will continue to iterate, and new competitors will constantly emerge. What truly keeps the development workflow stable and efficient is the intermediate management and orchestration layer. A mature AI Gateway can completely decouple model updates and iterations from the application layer, allowing developers to enjoy the dividends of models without repeatedly tinkering with configurations, managing scattered Keys, or manually handling failovers.
Developers interested in Qwen3.8-Max are advised to pay attention to two things simultaneously: one is the open-source weight release during the week of August 10, and the other is establishing an infrastructure capable of flexibly managing multi-model access. ServBay AI Gateway, as a full-featured AI gateway running locally, provides a relatively complete solution in protocol conversion, model mapping, channel management, and usage statistics, and can serve as a reference choice for building this infrastructure.