Ollama – v0.17.3: model: fix qwen3 tool calling in thinking (#14477)
🚨 Ollama v0.17.3 is live — and it’s fixing a big one for Qwen3 fans! 🎯
This patch (#14477) tackles a critical bug where Qwen3 and Qwen3-VL models were failing to properly handle tool calls during the “thinking” phase — i.e., before “ closes.
🔧 What’s fixed?
✅ Tool-call detection now works mid-think: The model correctly spots `<tool_call>` (tool call start tag) while still in thinking mode and smoothly transitions into tool-parsing — matching Hugging Face Transformers behavior.
✅ Robust tag parsing: Handles overlapping or partial tags (e.g., `<tool_call>` appearing before “) without breaking.
✅ Streaming-safe: Works reliably even when `<tool_call>` is split across chunks in streaming responses.
🧠 Why you’ll care:
This fix makes Qwen3-family models production-ready for agent workflows, tool-using assistants, and apps that rely on structured function/tool invocation — no more silent failures mid-call!
📦 Update now:
“`bash
ollama pull qwen3 # for text models
ollama pull qwen3vl # for vision-language variants
“`
Happy tool-calling! 🛠️✨