Tool calling, and which models have it
Check before you build. A model that answers prose beautifully will still fail the moment an agent expects a structured call.
The state of it
glm-5.3-xploded calls tools. Send a tools array and it comes back with finish_reason: tool_calls and arguments that parse, the same as any frontier model. That is the entire reason this model is the line: it is uncensored and it can drive an agent, and until recently nothing did both.
This page used to say the opposite, and it was right at the time. Uncensored models that could call tools were rare, so the only shape that worked was a mainstream model driving the session and an uncensored one called as one of its tools. That is no longer necessary and the guides no longer teach it.
Check it, do not assume it
Tool calling depends on how a model is served as well as on its weights, so a model card that says "supports function calling" is not evidence on its own. MagmaRouter reports supports_tools for every model it sells, and each model on the line was sent a real tool call before it was listed.
curl "https://api.magmarouter.com/v1/models?filter=xploded" -H "authorization: Bearer sk-mr-..."The call
Nothing here is specific to this gateway. It is an OpenAI-shaped tool call, and any client that already sends one needs no change.
curl "https://api.magmarouter.com/v1/chat/completions" -H "authorization: Bearer sk-mr-..." -H "content-type: application/json" -d '{
"model": "glm-5.3-xploded",
"messages": [{"role": "user", "content": "What is the weather in São Paulo?"}],
"tools": [{
"type": "function",
"function": {
"name": "get_weather",
"parameters": {
"type": "object",
"properties": {"city": {"type": "string"}},
"required": ["city"]
}
}
}]
}'The reply carries finish_reason: "tool_calls" and one call to get_weather with {"city": "São Paulo"}. Answer it with a role: "tool" message the way you would anywhere else.
It thinks before it answers
This is a reasoning model, and the thinking arrives in delta.reasoning ahead of any delta.content. On a short prompt that is a few hundred characters of reasoning before the first word of the reply, so a client that only renders content shows nothing for a second or two and then the whole answer at once.
That silence is the model working, not a stall. It matters for max_tokens: reasoning is counted as output, so a tight cap can be spent entirely on thinking and return an empty message with finish_reason: "length". Leave room, or leave the cap off.