1 Prerequisites

Before you start, make sure you have:

• 16 GB+ RAM (the 14B model needs ~10 GB VRAM or ~12 GB system RAM (varies by model size/quant; vendor guidance) with CPU offload)
• NVIDIA GPU with ≥ 12 GB VRAM recommended (RTX 3060 12 GB+, RTX 4060 Ti 16 GB, etc.) — CPU-only works but is slower
• ~15 GB free disk space for model weights
• Node.js 18+ — nodejs.org (or use Claude Code's native installer; Node is only needed for the npm install path)
• Claude Code (or Amp CLI) installed — Claude Code docs

2 Choose a Local Model Server

You need software that hosts Qwen locally and exposes an Anthropic-compatible /v1/messages API (the native protocol Claude Code speaks). Both options below now support it natively — no translation proxy needed:

This guide covers both paths. Pick one and follow the matching steps.

3a Path A — Install & Run with Ollama

1. Download & install from ollama.com/download
2. Open a terminal and pull the model. Qwen3-Coder is the current recommendation (30B-class MoE, ~3B active params, strong on agentic coding):

ollama pull qwen3-coder

Prefer the lighter, older Qwen 2.5 Coder 14B (smaller VRAM footprint)?

ollama pull qwen2.5-coder:14b

This downloads the Q4_K_M quantised weights (~9 GB). For the full-precision variant:

ollama pull qwen2.5-coder:14b-instruct-fp16

3. Verify it works — run a quick chat (use whichever tag you pulled):

ollama run qwen3-coder "Write hello world in Python"

4. Start the Ollama server (it usually auto-starts, but to be sure):

ollama serve

3b Path B — Install & Run with LM Studio

1. Download & install from lmstudio.ai
2. Open LM Studio → click Search → type qwen3-coder (or Qwen2.5-Coder-14B-Instruct for the lighter fallback)
3. Pick a quantisation (recommended: Q4_K_M for balanced speed/quality)
4. Click Download, wait for it to finish
5. Go to the Local Server tab (left sidebar, server icon)
6. Select the downloaded Qwen model from the dropdown
7. Click Start Server

4 Verify the Local API Works

Claude Code talks to local servers using the Anthropic Messages API (/v1/messages), not the OpenAI chat endpoint. Confirm your server exposes it:

For Ollama (port 11434):

curl http://localhost:11434/v1/messages \
  -H "Content-Type: application/json" \
  -H "anthropic-version: 2023-06-01" \
  -d '{
    "model": "qwen3-coder",
    "max_tokens": 256,
    "messages": [{"role":"user","content":"Say hello"}]
  }'

For LM Studio (port 1234):

curl http://localhost:1234/v1/messages \
  -H "Content-Type: application/json" \
  -H "anthropic-version: 2023-06-01" \
  -d '{
    "model": "qwen3-coder",
    "max_tokens": 256,
    "messages": [{"role":"user","content":"Say hello"}]
  }'

On Windows PowerShell, use Invoke-RestMethod or install curl via curl.se.

5 Configure Claude Code to Use Your Local Model

Point Claude Code at your local server's root URL (do not append /v1 - Claude Code adds the path itself). Set these before launching:

Linux / macOS (bash/zsh) - Ollama:

# For Ollama
export ANTHROPIC_BASE_URL=http://localhost:11434
export ANTHROPIC_AUTH_TOKEN=ollama
export ANTHROPIC_MODEL=qwen3-coder
export ANTHROPIC_DEFAULT_SONNET_MODEL=qwen3-coder
export ANTHROPIC_DEFAULT_HAIKU_MODEL=qwen3-coder
export ANTHROPIC_DEFAULT_OPUS_MODEL=qwen3-coder
export CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS=1
claude

Linux / macOS - LM Studio:

# For LM Studio
export ANTHROPIC_BASE_URL=http://localhost:1234
export ANTHROPIC_AUTH_TOKEN=lm-studio
export ANTHROPIC_MODEL=qwen3-coder
export ANTHROPIC_DEFAULT_SONNET_MODEL=qwen3-coder
export ANTHROPIC_DEFAULT_HAIKU_MODEL=qwen3-coder
export ANTHROPIC_DEFAULT_OPUS_MODEL=qwen3-coder
export CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS=1
claude

Windows PowerShell - Ollama:

$env:ANTHROPIC_BASE_URL = "http://localhost:11434"
$env:ANTHROPIC_AUTH_TOKEN = "ollama"
$env:ANTHROPIC_MODEL = "qwen3-coder"
$env:ANTHROPIC_DEFAULT_SONNET_MODEL = "qwen3-coder"
$env:ANTHROPIC_DEFAULT_HAIKU_MODEL = "qwen3-coder"
$env:ANTHROPIC_DEFAULT_OPUS_MODEL = "qwen3-coder"
$env:CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS = "1"
claude

Qwen 2.5 fallback: set ANTHROPIC_MODEL=qwen2.5-coder:14b (Ollama) or qwen2.5-coder-14b-instruct (LM Studio).

Why the tier vars? Claude Code internally requests claude-sonnet-... / claude-haiku-... / claude-opus-.... Mapping all three to your local tag prevents "model not found" rejections. CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS=1 strips Anthropic beta headers local servers do not understand.

6 Make the Configuration Persistent

Environment variables reset when you close the terminal. Persist them so you don't re-export every session:

Linux / macOS - add to ~/.bashrc or ~/.zshrc:

export ANTHROPIC_BASE_URL=http://localhost:11434
export ANTHROPIC_AUTH_TOKEN=ollama
export ANTHROPIC_MODEL=qwen3-coder

Windows PowerShell - set user-level variables (survives new shells):

[Environment]::SetEnvironmentVariable("ANTHROPIC_BASE_URL","http://localhost:11434","User")
[Environment]::SetEnvironmentVariable("ANTHROPIC_AUTH_TOKEN","ollama","User")
[Environment]::SetEnvironmentVariable("ANTHROPIC_MODEL","qwen3-coder","User")

To switch back to Anthropic's cloud, just unset the variables (Remove-Item Env:ANTHROPIC_BASE_URL on PowerShell, or unset ANTHROPIC_BASE_URL on bash) and relaunch claude.

?? Full Claude Code settings reference

7 Launch Claude Code

With your local server running and config in place, just start Claude Code normally:

claude

You should see it connect. Try a prompt:

You: Explain what a decorator is in Python, with an example.

Qwen3-Coder: A decorator is a function that takes another
function as an argument and extends its behavior without
modifying it...
    

8 Performance Tuning Tips

• GPU layers: In Ollama, set OLLAMA_NUM_GPU=999 to offload all layers to GPU
• Context window: Default is small. Raise it for code tasks (Qwen3-Coder supports a long context):

  ollama run qwen3-coder --num-ctx 32768

  Or in LM Studio: Settings → Context Length → 8192
• Quantisation trade-offs (vendor guidance; not independently benchmarked here):
  • Q4_K_M — best balance (speed + quality)
  • Q5_K_M — slightly better quality, needs more VRAM
  • fp16 — full precision, needs ~28 GB VRAM
• Batch size: If using CPU, increase batch size for faster prompt processing

📖 Ollama Modelfile options  |  Qwen Coder GitHub

9 Troubleshooting

10 Quick Reference Card