On September 1st, Tencent's AI assistant Marvis activated the 'custom model' feature. Users can now integrate third-party models that comply with generally accepted industry interface standards. These models include solutions from Tencent Cloud, Alibaba Cloud, DeepSeek, MiniMax, Kimi, Zhipu, and Xiaomi, as well as locally deployed open-source models. The configuration of these models is synchronized across different devices.
This move reflects an obvious shift in the large model market over the past year, as no single model remains optimal in all situations. Different models demonstrate superiority in long-text writing, code generation, or response speed and cost. Consequently, there has been a sharp increase in demand for the ability to switch between multiple models depending on the specific task requirements.
Previously, Marvis only offered platform-integrated models; now, users can combine a preview of Hunyuan Hy4 with DeepSeek-V4 Pro to create a single convenient preset, or direct the assistant to any model providing standard interfaces.
Access to this feature is available through the product's user interface: under 'Settings,' then 'Model Management.' Quick settings are available for Tencent Cloud, DeepSeek, and Alibaba Cloud, along with a special option for manually adding any required model. During a conversation, users can change models on the fly to match the current task.
Marvis provides every user with a daily free quota of 10 million tokens. Any usage through a connected third-party service is billed according to the user's own account and does not consume this free quota, which is beneficial for users who work intensively with code and long documents.
The update also lowers the requirements for local models. Marvis's previous 'local mode' required high computational power on the device. The new custom model feature allows open-source models to be run via a third-party application on one's own computer and then called in Marvis through simple setup. Since such conversation data is never transmitted to a cloud server, this is suitable for privacy-sensitive scenarios.
Finally, custom model settings are treated as account assets rather than device configurations. If a user logs into the same account on Windows, Mac, or mobile device, they retain their configured list of models, and the AI workflow is not interrupted when switching devices. Marvis already supports Mac, Windows, iOS, Android, and iPad, and this feature is available in the latest version.