Claude Code users chase local Qwen models

This is basically a small-model reality check: Claude Code’s agent loop seems to care less about model hype and more about tool-call reliability, and the community is already drifting upward in size.

• –One commenter recommends Qwen 3.5 4B, 9B, or 27B Unsloth quants and says to use the biggest model that fits.
• –Another says 9B is the floor, but 27B is meaningfully better, which suggests tiny 2B distills are too brittle for agentic coding.
• –The OP’s failed 2B distilled attempt reinforces a common pattern: underpowered local models often stumble on structured tool use before they fail at raw code generation.
• –Anthropic’s Claude Code docs center Anthropic-hosted model aliases, so local backends still look more like a workaround than an officially documented route.