Alibaba Cloud AI Gateway Supports DeepSeek V4

Alibaba Cloud AI Gateway has been updated to support the newly released DeepSeek-V4 models, enabling users to manage, call, and integrate these models...

Today, DeepSeek released two models, V4-Pro and V4-Flash. Their architecture and technical advantages can be summarized as follows:

Architecture upgrade: manifold-constrained residual connections mHC and the Muon optimizer improve training quality.
Post-training paradigm upgrade: introduces domain expert cultivation and on-policy distillation (fusing the capabilities of multiple experts into a single student model) to improve post-training efficiency and effectiveness.
Brand-new attention mechanism: introduces CSA and HCA on top of DSA to improve conversational performance under 1M context.

This brings improvements in model performance and cost-effectiveness, including:

Significantly improved Agent capabilities: In the Agentic Coding benchmark, V4-Pro has already reached the current best level among open-source models. According to evaluation feedback, its user experience is better than Sonnet 4.5, and its delivery quality is close to Opus 4.6 in non-thinking mode, but it still lags somewhat behind Opus 4.6 in thinking mode.
Rich world knowledge: DeepSeek-V4-Pro significantly outperforms other open-source models in world knowledge evaluations, falling only slightly short of the top closed-source model Gemini-Pro-3.1.
World-leading reasoning performance: In evaluations for math, STEM, and competitive coding, DeepSeek-V4-Pro surpasses all currently publicly benchmarked open-source models and achieves results comparable to the world’s top closed-source models.
With 1M token context, reasoning FLOPs are only 27% of the previous-generation V3.2, and KV cache usage is only 10%.

DeepSeek-V4 supports the OpenAI ChatCompletions interface and the Anthropic interface. When calling the new model API, the Model parameter needs to be changed to deepseek-v4-pro or deepseek-v4-flash.

Alibaba Cloud AI Gateway provides management capabilities for Model API, Agent API, and MCP Server, and now supports management of the DeepSeek-V4 API first. Through Alibaba Cloud AI Gateway, you can call DeepSeek-V4 API services, including thinking, multi-turn dialogue, Tool Call, Anthropic /v1/messages compatible calls, and more. It also supports integration of DeepSeek-V4 on Claude Code, and additionally implements fallback capabilities between DeepSeek-V4 and other models such as Qwen.

Configuration Method

Open the AI Gateway page, click to enter the console, and click the target instance ID. In the left navigation bar, click Model API, then click Create Model API.

After entering the Create Model API form, you can configure it as follows:

Protocol: Different protocols correspond to a set of built-in default routes for this scenario, used to quickly generate common OpenAI/DashScope/vLLM-compatible interfaces.
API Name: You can customize the API name. Note that the name must be globally unique (cannot be duplicated under the same account). English letters, numbers, underscore "_", and hyphen "-" are supported, and the name cannot exceed 64 characters.
Domain: The domain for accessing this API. Multiple domains are supported. The combination of the domain and the base path BasePath must be unique.
Base Path: Configure the API's base request path, which defaults to /. You can choose whether to enable remove when forwarding to backend services.

After configuration, run a test case:

Community

Alibaba Cloud AI Gateway Supports DeepSeek V4

Configuration Method

Read previous post:

Read next post:

Alibaba Cloud Native Community

You may also like

Comments

Alibaba Cloud Native Community

Related Products

Container Compute Service (ACS)

Container Service for Kubernetes

Tongyi Qianwen (Qwen)

Alibaba Cloud for Generative AI