On September 1, Marvis, the AI assistant under Tencent, launched the "Custom Model" feature, allowing users to freely integrate leading models from Tencent Cloud, Alibaba Cloud, DeepSeek, Minimax, Kimi, Zhipu, Xiaomi, and others, while also supporting open-source models deployed locally. This new feature removes the previous limitation of a single model, enabling users to switch between different models as needed. The platform provides a daily free quota of 10 million tokens; heavy users can access additional resources by integrating third-party services. Additionally, local model integration lowers the barrier to private deployment, ensuring user conversation data never passes through cloud servers, and custom model configurations can be synchronized across Windows, Mac, iOS, Android, and iPad devices.Author and source: AIBase
On September 1, Marvis, Tencent’s operating system-level AI assistant, officially launched its “Custom Model” feature.
Users can now freely integrate third-party models in Marvis that comply with industry-standard interfaces, including mainstream models provided and supported by vendors such as Tencent Cloud, Alibaba Cloud, DeepSeek, Minimax, Kimi, Zhipu, and Xiaomi. The platform also supports integration of locally deployed open-source models and enables seamless synchronization across devices.
"Brain swap" capability is now open—model selection is no longer limited.
Over the past year, the large model field has entered a fast-paced competition where a new leader emerges every month, making it difficult for a single model to remain optimal across all scenarios. Some models excel at long-form writing, others stand out in code generation, and some lead in response speed and cost-effectiveness. Users are increasingly demanding the ability to freely switch between multiple models.
The most noticeable change in this update is the increased flexibility in model selection. Previously, Marvis users could only use the platform’s built-in models; with the new feature, users can now choose between the pre-configured combinations of HunYuan Hy4preview and DeepSeek-V4Pro, or connect to other large models that comply with industry-standard interfaces.
From the product interface, users can access the “Custom Model” configuration option by navigating to “Settings — Model Management.” The system includes pre-configured quick options for models from Tencent Cloud, DeepSeek, Alibaba Cloud, and other providers, while also retaining the “Custom Model” entry for users to add their own desired models. After configuration, users can switch between different models at any time during conversations to select the most suitable one for their current task.
Notably, Marvis provides each user with a daily free allowance of 10 million tokens, sufficient for the vast majority of everyday use cases. However, under high-frequency or intensive usage scenarios such as code generation or processing lengthy documents, this allowance may be exhausted. By integrating custom models with third-party services, usage is counted toward the user’s own account and no longer consumes Marvis’s daily free quota, offering heavy users an unrestricted pathway.
Local Model Integration: Lowering the Barrier to Private Deployment
In addition to cloud-based models, this update now opens compatibility for local models. For users who prioritize data privacy or wish to reduce invocation costs, simply run open-source models on your own computer via a third-party application, then complete simple configuration in Marvis to access your local model as if using "Local Mode."
The practical value of this capability lies in lowering the barrier to entry. Previously, Marvis’s “Local Mode” required devices with substantial local computing power to run smoothly and had specific hardware requirements. By enabling users to connect custom models locally, they can now invoke local models in “Efficiency Mode,” thereby reducing the high performance demands previously imposed by “Local Mode.”
From a data security perspective, local model integration means users' conversation content and data do not pass through any cloud servers, making it ideal for scenarios with high data privacy requirements.
Cross-platform model asset synchronization: Configure once, use across all platforms
The third key feature is cross-device synchronization. Custom model configurations added by users are treated as personal account assets, tied to the account rather than a single device. This means that, when logging into Marvis on Windows, Mac, or mobile devices under the same account, previously configured model lists are automatically synchronized, eliminating the need to re-add them.
Marvis is now available on Mac, Windows, iOS, Android, and iPad. For professionals who frequently switch between devices, a model set up on your office computer today can be picked up right where you left off on your mobile device tomorrow while traveling—AI workflows are no longer interrupted by device boundaries.
Decoupling the AI assistant from the underlying model is becoming an industry-wide choice—users no longer need to switch entire apps to try a new model or toggle between multiple assistants. The custom model feature is now available in the latest version of Marvis; users can experience it after updating, under Settings → Model Management.