Optimizing Gemma 4 and Claude CLI for Macbook PRO M2,M3,M4,M5 Pro or Max

To leverage your Macbook PRO M2,M3,M4,M5 Pro or Max with the most optimized configuration for Gemma 4 and the Claude CLI, you should use a combination of persistent environment variables (for performance) and a custom model setup (for intelligence).

Here is the “final” script and configuration plan.

1. The “Optimized” Server Startup

Before running the Claude CLI, you need to start the Ollama server with Flash Attention and KV Cache Quantization enabled. This is what allows your 64GB Mac to handle 64k+ context smoothly.

Run this command to start your local server:

OLLAMA_FLASH_ATTENTION=1 OLLAMA_KV_CACHE_TYPE=q8_0 ollama serve

Tip: You can add these variables to your ~/.zshrc so you never have to type them again.

2. Create the “Claude-Optimized” Model

The Claude CLI works best when the model is specifically configured for high-context coding. Run these lines to create a specialized version of Gemma 4:

# 1. Create a temporary Modelfile
cat <<EOF > Gemma-Claude.Modelfile
FROM gemma4:31b-it-q8_0
PARAMETER num_ctx 65536
PARAMETER temperature 0.2
PARAMETER top_p 0.9
PARAMETER repeat_penalty 1.1
SYSTEM "You are an expert AI software engineer. You are acting as the backend for the Claude CLI. Follow all tool-use instructions precisely and keep your responses technical and concise."
EOF

# 2. Build the model in Ollama
ollama create gemma4-claude -f Gemma-Claude.Modelfile

# 3. Clean up the temp file
rm Gemma-Claude.Modelfile

3. Running the Session

Finally, you need to tell the Claude CLI to look at your local Mac instead of Anthropic’s servers.

Method A: The “Launch” Way (Ollama 2026 Native) without Step 3 If you haven’t set the environment variables, the newest versions of Ollama have a shortcut:

CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1 ollama launch claude --model gemma4-claude

Now, you have two ways to start coding:

Method B: The Direct Way

Add these lines to your ~/.zshrc (or just run them in your current terminal):

# Point Claude CLI to your local Ollama instance
export ANTHROPIC_BASE_URL="http://localhost:11434/v1"
export ANTHROPIC_API_KEY="ollama"
export ANTHROPIC_MODEL="gemma4-claude"

# Optional: Disable external telemetry for 100% local privacy
export CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1

Since you’ve set the environment variables above, simply type:

claude

Why this is the best script for you:

1. KV Cache Q8: This saves ~10-15GB of RAM compared to standard f16 cache, leaving more room for your IDE.
2. Flash Attention: Reduces the “thinking time” (prefill) when you send large files.
3. Gemma 4 31B Q8: You are using a high-precision 8-bit model, which is much more competent for coding than the standard 4-bit (q4) models most people use.
4. Temperature 0.2: Locked in via the Modelfile to ensure the CLI tools don’t hallucinate.

Final Check: In your terminal, run ollama ls. If you see gemma4-claude, you are ready to go!