Function Calling in Chinese Models: Support and Gaps

Technical background

“Function calling” and “tool calling” describe the same mechanism; the terminology split is historical. OpenAI introduced “function calling” in June 2023 in the Chat Completions API; Anthropic used “tool use”; the industry has since converged on “tools” as the umbrella term. Understanding both terms matters when reading vendor docs.

How it works

Functions are declared as JSON Schema definitions. When the model decides a function is needed, it returns a structured object: function name plus a JSON arguments payload. The application executes the function and sends the result back as a function/tool message, after which the model continues. This round trip is the unit of an agent loop.

Why it matters

Function calling converts language into structured, executable intent — the difference between a model that describes an action and one that triggers it. It underpins data extraction, database operations, and every agent framework.

Chinese adoption

In the China AI Hub database (last verified 2026-09-22), DeepSeek documents function calling most explicitly: DeepSeek-V4-Pro and V4.1-Flash both list the capability, and DeepSeek maintains dedicated function-calling documentation compatible with OpenAI-style schemas. Other Chinese providers implement the equivalent under “tools” — Qwen3.8-Max, Kimi K3, MiniMax M3 and Doubao Seed 2.1 Pro all list tool calling. The API layer confirms this is the practical surface: Ark, Model Studio, Kimi and Z.ai each expose OpenAI-compatible function/tool calling with structured output on at least their flagship models.

Major Chinese companies and models

• DeepSeek — V4-Pro and V4.1-Flash list function calling and structured output; official docs describe OpenAI-compatible invocation.

Practical applications

Natural-language interfaces to internal APIs, structured data extraction from documents, CRM/ERP automation, and compatibility layers for applications built on OpenAI-style tool schemas.

Limitations

JSON argument validity still needs validation (the model can emit malformed or wrong-typed arguments); schemas with many functions degrade selection accuracy; and nested or long argument payloads are error-prone.

Deployment considerations

Validate and coerce arguments before execution; keep function sets small per request; prefer enums and tight schemas; and version your function definitions since models are sensitive to description changes.

What the available evidence actually shows

The database records function/tool calling as a capability flag on the flagship models of all six tracked companies, and as an API surface on Ark, Model Studio, DeepSeek, Kimi and Z.ai — so the claim “function calling is broadly available across Chinese providers” is well evidenced. What is not evidenced is the reliability of schema adherence: no independent benchmark in the database measures structured-output validity or function-selection accuracy across these models. China AI Hub analysis indicates function calling is now a table-stakes compatibility layer — the meaningful differentiator has moved from “does it support it” to “how reliably does it emit valid arguments,” which the database cannot yet answer.

Future development

The distinction between function calling, tool use and MCP tools is collapsing into one capability surface; expect schema standards and tool registries to keep converging.

Labels used above: Official fact (from OpenAI’s and DeepSeek’s function-calling docs and the China AI Hub database), Vendor-reported claim (capability flags), and China AI Hub analysis (our synthesis, always introduced as such).