Marvis Custom Models: How Third-Party Model Support Is Supposed to Work

The clearest interpretation of Marvis’s reported custom-model feature is a bring-your-own-model (BYOM) or API configuration layer: users would enter a compatible service endpoint, API key, and model identifier, then select that model from the Marvis interface.

That is technically plausible. Tencent Cloud documents the exact ingredients required by a standard model integration: Bearer-token authentication, model identifiers returned by GET /v1/models, and compatibility with OpenAI-style APIs.1 However, the evidence supplied here does not confirm that Marvis launched this capability on September 1, 2026. It also does not verify the complete list of supported providers, the claimed settings path, local-runtime support, or synchronization across Windows, Mac, iOS, Android, and iPad.

What the proposed feature would do

A model-management system of this kind would let a user keep several model connections in one workspace rather than using only an application’s built-in models. The basic configuration would normally include:

• a provider or custom service;
• a base URL or API endpoint;
• an API key; and
• the provider’s model ID.

Tencent Cloud’s TokenHub documentation shows how this pattern works on the provider side. Its API uses an Authorization: Bearer <API KEY> header, and GET /v1/models can return currently available model IDs and their status.1 Tencent Cloud also documents OpenAI Chat Completions, OpenAI Responses, and Anthropic-compatible calling methods for listed models, although a model must be enabled in the platform before it can be called.4

That makes standard-interface support a reasonable technical description of the alleged Marvis feature. It is not, by itself, proof that Marvis has implemented the feature.

Can users switch between models during a conversation?

If Marvis accepts multiple compatible model configurations, users could theoretically select a different model for separate requests or conversations. For example, one request could use hy4-preview, while another could use a DeepSeek-V4-Pro model ID. Tencent Cloud lists both Hy4 preview and DeepSeek-V4-Pro among its available models.12

The word combine needs care here. The available evidence supports putting both models behind one client and switching between them. It does not establish that Marvis blends their weights, merges their internal reasoning, or automatically sends one model’s output to another. Unless Marvis documents an orchestration or routing feature, “combining Hy4 preview and DeepSeek-V4 Pro” should be read as making both available in one workflow, not creating a new hybrid model.

What is verified about Hy4 preview and DeepSeek-V4 Pro?

Tencent Cloud lists Hy4 preview with the model ID hy4-preview, a 1-million-token context window, and capabilities including deep reasoning, structured output, function calling, and caching.2 Tencent’s release materials also describe Hy4 preview as an open-source model available through Tencent Cloud TokenHub and OpenRouter.12

The same Tencent API documentation lists several DeepSeek-V4-Pro identifiers and shows that the model is available through compatible API protocols.14 DeepSeek separately announced the release of its V4 Pro model through its app, website, and API.14

These sources confirm that the two models have compatible access paths in at least some provider environments. They do not confirm a Marvis-specific Hy4-plus-DeepSeek workflow.

Which providers does Marvis support?

The claim names Tencent Cloud, Alibaba Cloud, DeepSeek, MiniMax, Kimi, Zhipu GLM, Xiaomi, and locally deployed open-source models. The supplied sources show that several of these providers or model families exist in broader model-platform ecosystems. For example, Tencent describes TokenHub as aggregating Tencent, DeepSeek, MiniMax, Kimi, Zhipu GLM, and Qwen models.26 Alibaba Cloud’s model documentation also lists DeepSeek and Kimi model identifiers.22

But that is different from proving direct Marvis support. One Marvis-focused source explicitly says that Marvis did not directly support user-supplied third-party model APIs at the time it was published, instead describing built-in cloud and local modes.20 Other sources describe Marvis’s built-in Hunyuan, DeepSeek, and Qwen-based arrangements, rather than a user-configurable provider marketplace.2528

Bottom line: the full Marvis provider list remains unverified from the evidence provided. Users should not assume that a model being available through Tencent Cloud, Alibaba Cloud, or another platform means it can be added directly to Marvis.

Would third-party usage use Marvis’s free quota?

Marvis was described by Tencent-related materials as offering each user 10 million free tokens per day for its existing service.4749 A separately configured provider API would normally authenticate requests with the user’s own API key, meaning the provider—not the client interface—would control access, usage limits, and billing.

That is how Tencent Cloud’s API and TokenHub systems are documented: API keys authorize calls, and model access depends on the user’s enabled services or plan.1431 Still, the supplied sources do not contain an official Marvis statement confirming that custom-model requests would be excluded from Marvis’s daily allowance or billed entirely to the user’s external account. That accounting rule should therefore be treated as an expected implementation detail, not a verified Marvis policy.

What about local open-source models and privacy?

A locally deployed model can run inference on a computer or private server instead of sending prompts to a public cloud model provider. Tencent documentation separately confirms that its services can support conversations with self-deployed open-source models.43 In principle, using a local endpoint can reduce the amount of conversation data sent to an external inference service.

The privacy conclusion depends on the actual architecture, however. The evidence does not identify Marvis’s supported local engines, endpoint requirements, networking behavior, or whether Marvis sends telemetry, files, prompts, or intermediate task data to its own servers. It therefore cannot substantiate a formal claim that a Marvis local-model session is guaranteed to remain entirely on-device.

This distinction matters: local inference is a privacy property of the deployment, while Marvis’s privacy behavior is a product-level claim that requires Marvis documentation or testing.

Is model configuration synchronized across devices?

The evidence describes Marvis as a cross-device product and reports support for multiple platforms, but it does not document synchronization of custom model definitions, API keys, endpoints, or model preferences.4750 Those are sensitive settings, and synchronization would raise separate questions about encryption, account storage, and whether credentials are copied between devices.

Accordingly, the claim that a model setup automatically syncs across Windows, Mac, iOS, Android, and iPad cannot be confirmed from the provided sources. Even if Marvis supports cross-device account access, that does not necessarily mean custom endpoints or API credentials follow the user.

The practical verdict

Marvis custom-model support is best described as an unverified BYOM/API concept, not a fully documented launch feature. The provider-side technology is real: standard API authentication, model IDs, compatible protocols, cloud-hosted models, and self-deployed inference are all documented in the surrounding ecosystem.142643

What remains unproven is the Marvis-specific layer:

• whether the feature launched on September 1, 2026;
• whether Settings > Model Management is the actual configuration path;
• which third-party providers are supported directly;
• whether external calls use the user’s own billing account;
• which local runtimes are compatible; and
• whether model settings and credentials synchronize across devices.

Until Tencent publishes Marvis documentation covering those points, the safest conclusion is that Marvis has documented built-in cloud and local operating modes, while the broader third-party model-switching workflow remains a claim that still needs primary confirmation.