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

Description and use

Local LLMs can reliably act as coding agents when the right model family (Qwen 3.x, GLM‑4.x, Devstral Small 2) is chosen and the environment is configured for large context windows; a single 24GB GPU running Qwen 3.6 27B is the sweet spot for most users.

Problem

Local LLMs claim to match Claude Code on SWE‑bench, but the benchmark does not measure the reliability needed for agentic coding workflows.

Outcome

Local LLMs can reliably act as coding agents when the right model family (Qwen 3.x, GLM‑4.x, Devstral Small 2) is chosen and the environment is configured for large context windows; a single 24GB GPU running Qwen 3.6 27B is the sweet spot for most users.

Workflow

• 1. Pull a capable coder model (e.g., qwen3-coder:30b) via Ollama.
• 2. Set environment variables to point Claude Code to the local endpoint: ANTHROPIC_BASE_URL="http://localhost:11434" and ANTHROPIC_AUTH_TOKEN="local-dummy".
• 3. Run Claude Code with the local model.
• 4. For Codex CLI, use the --oss flag to point to the local Ollama endpoint or configure LM Studio’s config.toml.
• 5. Explicitly set a large context window (32K–64K) to avoid early starvation.
• 6. Verify tool‑call reliability by testing multi‑step agentic scenarios.
• 7. Adjust GPU choice based on hardware tier: 24GB for Qwen 3.6 27B, 8GB for GLM‑4.7‑Flash, etc.

Evidence

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

Source

Tool‑call validity — can it emit a well‑formed edit_file / run_command call the harness can parse, *reliably*, not usually?

Source

The curve climbs steeply up to ~24GB and then **flattens** — the 480B coder even dips *below* the 27B on Verified while costing 3× the memory.

Source