Can a local LLM replace Claude Code? The benchmark everyone quotes is the wrong one.

说明与用途

本地LLM可以在正确选择模型、匹配硬件并验证工具调用可靠性的前提下替代Claude Code进行编码任务，但需要仔细配置和监控以避免代理失败。

问题

Local LLMs can write correct functions but fail to reliably act as agents that emit valid tool calls, maintain long‑horizon consistency, handle large context, and follow instructions, making them unsuitable for real‑world editor integration.

结果

Local LLMs can replace Claude Code for coding tasks when the correct model family is chosen, hardware is matched, and tool‑calling reliability is verified, but they require careful configuration and monitoring to avoid agentic failures.

工作流

• Identify hardware constraints and desired agentic capabilities.
• Select an appropriate model family (e.g., Qwen 3.6 27B for 24GB GPU, GLM‑4.7‑Flash for 8GB laptop).
• Pull the model via Ollama (e.g., ollama pull qwen3-coder:30b).
• Set environment variables: export ANTHROPIC_BASE_URL="http://localhost:11434" and export ANTHROPIC_AUTH_TOKEN="local-dummy".
• Run Claude Code (claude) or Codex CLI with --oss pointing to the local server.
• Ensure context window is set to at least 32K (preferably 64K) to avoid early starvation.
• Monitor tool‑call reliability and retry behavior; adjust model size if frequent malformed calls occur.
• Unset ANTHROPIC_BASE_URL to revert to cloud if needed.

证据

SWE‑bench measures "given a bug, produce a correct patch." That's a *writing* task.

来源

Tool‑calling is a separate trained skill from writing code.

来源

Reliability falls off a cliff below ~4B parameters.

来源

Set the context window explicitly — at least 32K, ideally 64K+.

来源