Qwen3-Coder-Next-GGUF Chats, Claude Code Rejects Tools

Strong signal for local-LLM users trying to wire open models into agentic coding tools, but this is really a tool-calling/runtime compatibility issue, not a general “model is broken” report.

• –Claude Code expects a model endpoint that advertises and handles tool use; a plain GGUF chat model can work interactively yet still be rejected for agent workflows.
• –The Hugging Face model card for `unsloth/Qwen3-Coder-Next-GGUF` recommends more than 45 GB unified memory for 4-bit quants, so an RTX 3060 12 GB is not a good fit for this 80B-class model.
• –The post is useful as a cautionary example: local model selection, quantization, and tool-call support all matter separately.
• –The privacy questions are relevant but ancillary; they turn the thread into a broader “how private is Claude Code with local models?” discussion.
• –Best editorial framing: “local model works in chat, but not every coding agent can use it as a tool-enabled backend.”