Chinese LLM Tool Calling Compatibility: A Systematic Comparison (as of Aug 2026)

Chinese providers of large language models (LLMs) have rapidly developed their tools to support OpenAI-style function calling. As of August 2026, five major LLM families from China — DeepSeek, GLM, Qwen, Kimi, and MiniMax, comprising 10 production variants — all provide OpenAI-compatible tool-calling endpoints. While swapping the API base URL generally yields valid responses, compatibility is not a simple binary switch.

Variations in payload schema, parallel tool-call behavior, streaming of tool-call data, and error handling still differ significantly enough that moving from one model to another may require detailed adjustments rather than a straightforward configuration change.

Testing across all variants reveals that while every model supports an OpenAI-compatible `/chat/completions` endpoint, compatibility exists on a spectrum. Among the tested models, GLM-5.1 and MiniMax M3 achieve the highest scores of 70.1 on the BenchLM tool-use suite, surpassing GPT-5.5's score of 67.8 as of August 21, 2026. Qwen3.7-Max follows closely with a score of 68.8 but is accessible only through Alibaba Cloud Model Studio within mainland China.

Kimi K2 serves as an open-weight option, yet its K2.6 iteration scores 60.5 on the same benchmark and has a history of non-standard streaming tool-call deltas. Router services like OpenRouter, Requesty, and Eden AI can normalize the transport layer but cannot address semantic or payload differences.

The evaluation methodology involved a unified harness that assessed all Chinese LLMs with identical function schemas and prompts using mock services. The OpenAI GPT-5.5 model served as the baseline for comparison. Additionally, three router layers were examined to quantify how much they normalize discrepancies in protocol and semantics.

Six key dimensions were evaluated: overall API protocol compatibility, fidelity of JSON payload schemas, parallel support for tool calls, streaming of tool-call deltas, error recovery mechanisms, and overall integration effort.

The landscape of deployment for these models can be categorized into three tiers.

1. Native OpenAI-compatible models include DeepSeek-V4-Pro, GLM-5.3, and the Qwen3.x series. These hosted APIs accept standard function schemas and offer options like 1 million-token context windows with adjustable reasoning efforts.

2. Self-hostable open-weight models comprise the Kimi K2 family, a 1 trillion-parameter mixture-of-experts model with 32 billion active parameters. This model is designed for coding, tool calling, and agentic tasks and can be deployed using various runtime environments such as vLLM, SGLang, KTransformers, or TensorRT-LLM.

3. Closed API-only models consist of Qwen3.7-Max, hosted exclusively on Alibaba Cloud in mainland China, which presents data residency implications for enterprises in Europe or the United States.

Specific provider notes reveal that MiniMax M3, despite being OpenAI-compatible for basic calls, introduces unique extension fields requiring additional parsing. Similarly, Kimi K2.6 accepts standard payloads but historically generated non-standard streaming tool-call deltas, necessitating version-specific parsers. Router services normalize the transport protocol but do not eliminate these semantic compatibility issues.

In terms of performance, BenchLM tool-use benchmark scores indicate that Chinese LLMs have outperformed GPT-5.5 at the top of the chart. GLM-5.1 and MiniMax M3 achieved scores around 70.1, followed by Qwen3.7-Max at 68.8, with GPT-5.5 scoring 67.8. Kimi K2.6 trailed with a score of 60.5. For long-running multi-step orchestration, Alibaba demonstrated Qwen3.7-Max maintaining continuous operation for 35 hours with 1,158 cumulative tool calls, showcasing its strong context retention.

However, availability is limited; Qwen3.7-Max is only accessible via Alibaba Cloud Model Studio in mainland China, while Kimi K2's open weights offer greater flexibility for self-hosted agentic applications.

Latency and throughput were critical during the unified-harness tests, with parser stability often emerging as a significant challenge due to varying streaming formats of tool-call deltas. Error recovery also varied significantly, with some models retrying or silently failing upon encountering malformed JSON, downstream exceptions, or empty results. Router layers can normalize the transport layer but cannot address semantic differences or ensure consistent error handling across providers.