minicpm5-2b native tool calls never parse

What is the issue?

With tools passed to /api/chat (and /v1/chat/completions), MiniCPM5-2B attempts the native XML tool call, but the response always has tool_calls: null (native) / missing (OpenAI compat) and content contains a mangled fragment such as name="get_weather"> name="city">Paris. The model reasoning (thinking/reasoning) shows correct tool intent. Standard clients that branch on tool_calls or finish_reason: tool_calls never take the tool branch. finish_reason/done_reason is always stop.

repro:

curl http://localhost:11434/api/chat -H "Content-Type: application/json" -d '{
  "model": "minicpm5-2b:latest",
  "messages": [{"role": "user", "content": "What is the weather in Paris? Use the tool."}],
  "stream": false,
  "tools": [{"type": "function", "function": {"name": "get_weather",
    "description": "Get weather for a city",
    "parameters": {"type": "object",
      "properties": {"city": {"type": "string"}},
      "required": ["city"]}}}]
}'

response:

{"model":"minicpm5-2b:latest",
 "message":{"role":"assistant",
  "content":" name=\"get_weather\"> name=\"city\">Paris",
  "thinking":"The user wants to know the weather in Paris. I need to use the get_weather function.\n"},
 "done":true,"done_reason":"stop"}

tool_calls is null. OpenAI compat (/v1/chat/completions) returns the same mangled content, a correct reasoning field, no tool_calls, finish_reason: stop.

upstream reference:

• vLLM minicpm5xml_tool_parser.py: "Tool XML tags are special tokens in MiniCPM5; must not strip them before tool parsing ... request.skip_special_tokens = False."
• SGLang documents --tool-call-parser minicpm5 for MiniCPM5-1B XML tool calls.
• OpenBMB/MiniCPM states SGLang is the recommended backend for MiniCPM5 tool calling.

Relevant log output

OS

Windows

GPU

Nvidia

CPU

Intel

Ollama version

0.30.8