DeepSeek‑R1‑0528 Model Gains Function‑Calling, JSON Output, and MCP Integration

DeepSeek recently released a minor version upgrade of its R1 large model, now labeled DeepSeek‑R1‑0528 . The upgrade claims substantial improvements in complex logical reasoning, long‑text stability, and code‑generation quality, with official statements highlighting benchmark performance gains, enhanced front‑end capabilities, reduced hallucination, and support for JSON output and function calling.

Function‑Calling Capability

The model can recognize when a response requires an external function, generate the appropriate parameters, invoke the function, receive the result, and continue generating a complete answer. This mechanism extends the model’s knowledge beyond its internal data, improving task efficiency and accuracy.

Enhanced flexibility: The model can dynamically call external tools or services based on task needs.

Improved response accuracy: Access to up‑to‑date external data reduces errors caused by knowledge limits.

Simplified integration via MCP: The Model Context Protocol (MCP) provides a standardized communication protocol, making it easier to connect the model with various external systems.

Knowledge Distillation and the 8B Model

Alongside the R1‑0528 release, DeepSeek announced an 8B model distilled from Qwen‑3. The distilled model claims slightly better performance than Qwen‑3‑235B while being far smaller. Knowledge distillation is described using a teacher‑student analogy: a large, powerful “teacher” model transfers its knowledge to a lightweight “student” model, which learns from the teacher’s output distributions to achieve comparable performance with reduced computational cost.

This approach is valuable for deploying AI models on resource‑limited devices such as smartphones or edge hardware, where the large model’s resource demands are prohibitive.

Integrating DeepSeek‑R1‑0528 with MCP

MCP (Model Context Protocol), introduced by Anthropic, standardizes the integration of large language models with external tools and data sources. Using MCP, developers can expose services such as train‑ticket queries or system‑time retrieval to the model.

Example MCP server configuration (JSON):

{
  "mcpServers": {
    "tongchenglvxing-mcp-server": {
      "command": "npx",
      "args": ["-y", "@wuchubuzai/tongchenglvxing-mcp-server"]
    }
  }
}

The server provides two tools: query_train_tickets_list: Retrieves train‑ticket information from Tongcheng Travel based on departure station, arrival station, and travel date.

System‑time query tool.

Parameters for query_train_tickets_list include depStationName (string), arrStationName (string), and depDate (string, format yyyy‑MM‑dd).

To use the model, a third‑party proxy service provides access to DeepSeek‑R1‑0528. After registering at the provided URL, users configure the proxy address and API key in their AI client, then add the MCP server to Cherry Studio (or a similar tool). The model can then invoke the MCP services during inference.

Security and Limitations

While MCP offers powerful integration, enterprise deployment requires security controls. A recent software conference presentation by Lenovo’s Li Bin highlighted the need for mechanisms to mitigate risks when exposing external tools via MCP.

Li Bin also discussed MCP’s limitations, emphasizing careful governance to avoid security vulnerabilities.

Conclusion

The DeepSeek‑R1‑0528 update delivers notable performance and functional enhancements, particularly function calling and JSON output, which enable practical integration with MCP services. Combined with the newly released 8B distilled model, developers gain flexible, resource‑efficient options for deploying advanced LLM capabilities in real‑world applications.