Tencent's Marvis Adds Third-Party AI Model Support
In Marvis's Settings, under Model Management, a new custom-model entry lets users add an AI service that follows common industry interface standards. Tencent's operating-system-level assistant can now call models from Tencent Cloud, Alibaba Cloud, DeepSeek, MiniMax, Kimi, Zhipu and Xiaomi, alongside models deployed locally by the user. Pandaily reports that the feature went live in the latest version on September 1, 2026.
The reason is straightforward: no single model is consistently best at every task. Some are stronger at long-text writing, others at code generation, while response speed and cost can decide a different choice. Marvis users can switch models during a conversation, or pair the bundled Hunyuan Hy4 preview with DeepSeek-V4 Pro in one preset instead of managing each task in a separate assistant.
The update also changes what “local” means in practice. Marvis's earlier local mode required substantial on-device computing power; the custom-model route lets users run an open-source model through a third-party app on their own computer and connect it with a simple configuration. Because the conversation data does not touch a cloud server, the setup is suited to privacy-sensitive scenarios, though the model itself still has to be operated on the user's computer.
There is a billing boundary. Marvis gives every user a daily free quota of 10 million tokens, but calls made through a connected third-party service are charged to the user's own account and do not consume that allowance. That distinction matters most for intensive code generation and long-document processing, where usage can accumulate quickly.
So what changes concretely? A Marvis user moving between Windows, Mac or mobile no longer has to rebuild the model list: custom-model settings are stored as account assets and sync across devices. The result is less dependence on one bundled model and a simpler path to private, locally run options, while usage through each connected service is billed to the user's account.
Comments
Loading the thread…
Sign in to leave a comment. Sign in