Michael Limberger

The engine, then a model that can actually see

Ollama runs your local models. Install it, pull a model, then create a properly configured version of that model. That last step is critical and often missed.

+-------------------------------------------------------+
|  IMPORTANT: Ollama 0.14.0 or higher required          |
|                                                       |
|  The Anthropic API compatibility was added January    |
|  2026. Older versions return 404 errors.              |
+-------------------------------------------------------+

Show me. Check version: ollama --version. If it is below 0.14.0, update.

Fresh install on macOS: download from ollama.com, or run brew install ollama. On Linux: curl -fsSL https://ollama.ai/install.sh | sh.

Already have it but too old? On macOS Homebrew: brew upgrade ollama. Direct install: download the latest from ollama.com and install over the old version. On Linux, the same install script handles upgrades.

Verify: ollama --version should show 0.14.0 or higher. On macOS, open Ollama from Applications (menu bar icon) or run ollama serve. On Linux: sudo systemctl start ollama, then sudo systemctl enable ollama so it starts on boot.

Pull the base, then don't use the default context

We recommend qwen3-coder-next:

ollama pull qwen3-coder-next:latest

About 18GB, so go get coffee.

+-------------------------------------------------------+
|  DO NOT SKIP THIS SECTION                             |
|                                                       |
|  The base model defaults to 4K context.               |
|  Claude Code will barely function with 4K.            |
|  You MUST create a model with proper context size.    |
+-------------------------------------------------------+

qwen3-coder-next supports 256,000 tokens of context. Ollama's default is 4,096. Ollama uses a conservative default regardless of model capability. With only 4K, Claude Code cannot fit your files plus conversation history. It will seem broken.

Show me. Create a Modelfile at ~/Modelfile-qwen-claude:

FROM qwen3-coder-next:latest
PARAMETER num_ctx 32768

FROM uses qwen3-coder-next as the base. num_ctx sets the context window to 32K tokens. Then:

ollama create qwen3-coder-32k -f ~/Modelfile-qwen-claude

This creates qwen3-coder-32k: identical to the base, but with 32K context instead of 4K. It does not re-download the weights. It creates a configuration layer. ollama list should show both the 18GB base and the 18GB 32k version. Use the 32k version with Claude Code.

Test: ollama run qwen3-coder-32k, then in another terminal run ollama ps. CONTEXT should show 32768, not 4096. Type /bye to exit.

Pick a window that fits RAM

On 16GB unified memory: PARAMETER num_ctx 8192, and name it qwen3-coder-8k. On 32GB: 32768, qwen3-coder-32k. On 64GB: 65536, qwen3-coder-64k. Larger context uses more RAM. Do not exceed available memory or the machine swaps to disk and slows to a crawl.

If the client is on another machine

By default Ollama only listens on localhost. Skip this section if everything is on one machine.

On macOS: echo 'export OLLAMA_HOST="0.0.0.0:11434"' >> ~/.zshrc, then source ~/.zshrc, pkill ollama, and ollama serve. That runs in the foreground. Keep the terminal open.

On Linux: sudo systemctl edit ollama.service, add a [Service] section with Environment="OLLAMA_HOST=0.0.0.0:11434", then daemon-reload and restart. Check lsof -i :11434. You want *:11434 (LISTEN). An asterisk means all interfaces. If it says localhost, the setting did not take.

Find the server IP. On macOS: ipconfig getifaddr en0. On Linux: hostname -I | awk '{print $1}'. Write it down.

Show me. Test the Anthropic endpoint Claude Code will use. Use your configured model name, not the base:

curl http://YOUR_IP:11434/v1/messages \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen3-coder-32k",
    "max_tokens": 50,
    "messages": [{"role": "user", "content": "Say hello"}]
  }'

You should get JSON back with a greeting.

When Ollama won't talk

Connection refused: is Ollama running (ps aux | grep ollama)? Is the firewall blocking 11434? Did OLLAMA_HOST take effect (lsof -i :11434)? "404 page not found" on /v1/messages: the version is too old, so update to 0.14.0+. Model not found: did you create the configured model, and is the name spelled exactly as in ollama list?

Next: install Claude Code on the workstation.