{"id":3719,"date":"2026-06-18T05:49:28","date_gmt":"2026-06-18T05:49:28","guid":{"rendered":"https:\/\/aiopsschool.com\/blog\/?p=3719"},"modified":"2026-06-18T05:49:30","modified_gmt":"2026-06-18T05:49:30","slug":"how-to-build-a-self-hosted-qwen-ai-coding-cluster-using-multiple-laptops","status":"publish","type":"post","link":"https:\/\/aiopsschool.com\/blog\/how-to-build-a-self-hosted-qwen-ai-coding-cluster-using-multiple-laptops\/","title":{"rendered":"How to Build a Self-Hosted Qwen AI Coding Cluster Using Multiple Laptops"},"content":{"rendered":"\n

<\/p>\n\n\n\n

    \n
  1. Self-Hosted Qwen Coding Cluster: Run AI Coding Agents Across MacBook and Intel Laptops<\/strong><\/li>\n\n\n\n
  2. Build Your Own Local AI Coding Cluster with Qwen, Ollama, LiteLLM, and Qwen Code<\/strong><\/li>\n\n\n\n
  3. How to Use Multiple Laptops as a Self-Hosted AI Coding Cluster<\/strong><\/li>\n\n\n\n
  4. Qwen Coding Cluster Tutorial: Pool MacBook and Intel Laptops for Local AI Development<\/strong><\/li>\n\n\n\n
  5. From Spare Laptops to AI Coding Farm: Self-Hosting Qwen for Software Development<\/strong><\/li>\n\n\n\n
  6. Complete Guide to Building a Local Qwen AI Coding Agent Cluster<\/strong><\/li>\n\n\n\n
  7. Run Qwen Locally: Multi-Laptop AI Coding Cluster Setup for Developers<\/strong><\/li>\n\n\n\n
  8. Self-Hosted AI Coding Farm with Qwen: MacBook M4 + Intel Laptop Setup<\/strong><\/li>\n\n\n\n
  9. How to Create a Private AI Coding Cluster Using Qwen and Ollama<\/strong><\/li>\n\n\n\n
  10. Build a Local Coding Agent Cluster with Qwen: Step-by-Step Developer Guide<\/strong><\/li>\n<\/ol>\n\n\n\n

    <\/p>\n\n\n\n

    3\u20134 laptops \u2260 one single pooled Qwen brain\n3\u20134 laptops = multiple local Qwen model servers + one router + multiple coding agents\n<\/code><\/pre>\n\n\n\n
    Below is a complete practical tutorial.<\/p>\n\n\n\n
    \n\n\n\n
    Self-Hosted Qwen Coding Cluster Using 3\u20134 Laptops<\/h1>\n\n\n\n
    0. What we are building<\/h2>\n\n\n\n
    You will have:<\/p>\n\n\n\n
    Controller Laptop\n  - LiteLLM router\n  - Qwen Code CLI\n  - Git worktrees\n  - task runner scripts\n\nWorker Laptop 1\n  - Ollama\n  - Qwen model\n\nWorker Laptop 2\n  - Ollama\n  - Qwen model\n\nWorker Laptop 3\n  - Ollama\n  - Qwen model\n<\/code><\/pre>\n\n\n\n
    Final flow:<\/p>\n\n\n\n
    Qwen Code \/ Aider \/ VS Code\n        \u2193\nLiteLLM Router :4000\n        \u2193\n \u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510\n \u2502 m4-1.local:11434      \u2502 m4-2.local:11434      \u2502 intel-1.local:11434   \u2502\n \u2502 qwen3-coder:30b       \u2502 qwen3-coder:30b       \u2502 qwen2.5-coder:14b\/32b \u2502\n \u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2534\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2534\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518\n<\/code><\/pre>\n\n\n\n
    This gives you a local AI coding cluster<\/strong>, not a fake RAM\/CPU pooling trick.<\/p>\n\n\n\n
    \n\n\n\n
    1. Recommended model selection<\/h1>\n\n\n\n
    For your hardware, do not start with Qwen 480B. Ollama lists qwen3-coder:480b<\/code> as a local model requiring at least 250GB memory\/unified memory<\/strong>, so your 24GB MacBooks are not the right machines for that. The practical target is qwen3-coder:30b<\/code> or qwen2.5-coder:14b<\/code>. Ollama lists qwen3-coder:30b<\/code> as around 19GB<\/strong> with a 256K context window, while qwen2.5-coder:14b<\/code> is around 9GB<\/strong> with 32K context. (ollama.com<\/a>)<\/p>\n\n\n\n
    Use this:<\/p>\n\n\n\n
    Machine<\/th>First model to try<\/th>Safer fallback<\/th><\/tr><\/thead>
    M4 MacBook Pro 24GB #1<\/td>qwen3-coder:30b<\/code><\/td>qwen2.5-coder:14b<\/code><\/td><\/tr>
    M4 MacBook Pro 24GB #2<\/td>qwen3-coder:30b<\/code><\/td>qwen2.5-coder:14b<\/code><\/td><\/tr>
    Intel i7 64GB<\/td>qwen2.5-coder:32b<\/code> only if patient<\/td>qwen2.5-coder:14b<\/code><\/td><\/tr>
    Extra laptop<\/td>same as above<\/td>qwen2.5-coder:7b<\/code><\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n
    Important: 24GB Mac can load a 19GB model, but context\/KV cache also needs memory. Ollama says larger context increases memory usage, and parallel requests multiply context memory usage. For coding agents on 24GB machines, start with 8K or 16K context<\/strong>, not 64K\/256K. (docs.ollama.com<\/a>)<\/p>\n\n\n\n
    \n\n\n\n
    2. Name your laptops<\/h1>\n\n\n\n
    Use simple hostnames.<\/p>\n\n\n\n
    Example:<\/p>\n\n\n\n
    m4-1.local\nm4-2.local\nintel-1.local\ncontroller.local\n<\/code><\/pre>\n\n\n\n
    On macOS<\/h2>\n\n\n\n
    Run this on each Mac, changing the name:<\/p>\n\n\n\n
    sudo scutil --set HostName m4-1\nsudo scutil --set LocalHostName m4-1\nsudo scutil --set ComputerName m4-1\ndscacheutil -flushcache\nhostname\n<\/code><\/pre>\n\n\n\n
    For the second Mac:<\/p>\n\n\n\n
    sudo scutil --set HostName m4-2\nsudo scutil --set LocalHostName m4-2\nsudo scutil --set ComputerName m4-2\ndscacheutil -flushcache\nhostname\n<\/code><\/pre>\n\n\n\n
    On Ubuntu\/Linux Intel laptop<\/h2>\n\n\n\n
    sudo hostnamectl set-hostname intel-1\nsudo apt update\nsudo apt install -y avahi-daemon\nsudo systemctl enable --now avahi-daemon\nhostname\n<\/code><\/pre>\n\n\n\n
    Now test from your controller:<\/p>\n\n\n\n
    ping m4-1.local\nping m4-2.local\nping intel-1.local\n<\/code><\/pre>\n\n\n\n
    \n\n\n\n
    3. Install Ollama on every worker laptop<\/h1>\n\n\n\n
    Ollama supports Apple M-series Macs with CPU\/GPU support, while x86 Macs are CPU-only according to its macOS requirements. On Linux, Ollama can run as a systemd service, and NVIDIA\/AMD GPU setup is optional depending on hardware. (docs.ollama.com<\/a>)<\/p>\n\n\n\n
    macOS workers<\/h2>\n\n\n\n
    Install Ollama:<\/p>\n\n\n\n
    curl -fsSL https:\/\/ollama.com\/install.sh | sh\n<\/code><\/pre>\n\n\n\n
    Verify:<\/p>\n\n\n\n
    ollama --version\nollama list\n<\/code><\/pre>\n\n\n\n
    Ubuntu\/Linux worker<\/h2>\n\n\n\n
    curl -fsSL https:\/\/ollama.com\/install.sh | sh\nsudo systemctl status ollama\nollama --version\n<\/code><\/pre>\n\n\n\n
    \n\n\n\n
    4. Configure Ollama for LAN cluster mode<\/h1>\n\n\n\n
    By default, Ollama binds to 127.0.0.1:11434<\/code>. To expose it to your LAN, set OLLAMA_HOST=0.0.0.0:11434<\/code>. Ollama\u2019s docs specifically say to use launchctl<\/code> on macOS and systemd environment overrides on Linux. (docs.ollama.com<\/a>)<\/p>\n\n\n\n
    On macOS workers<\/h2>\n\n\n\n
    Run this on each Mac:<\/p>\n\n\n\n
    launchctl setenv OLLAMA_HOST \"0.0.0.0:11434\"\nlaunchctl setenv OLLAMA_CONTEXT_LENGTH \"8192\"\nlaunchctl setenv OLLAMA_NUM_PARALLEL \"1\"\nlaunchctl setenv OLLAMA_MAX_LOADED_MODELS \"1\"\nlaunchctl setenv OLLAMA_KEEP_ALIVE \"30m\"\nlaunchctl setenv OLLAMA_NO_CLOUD \"1\"\n<\/code><\/pre>\n\n\n\n
    Restart Ollama:<\/p>\n\n\n\n
    osascript -e 'quit app \"Ollama\"' || true\nopen -a Ollama\n<\/code><\/pre>\n\n\n\n
    Check:<\/p>\n\n\n\n
    curl http:\/\/localhost:11434\/api\/tags\n<\/code><\/pre>\n\n\n\n
    On Linux worker<\/h2>\n\n\n\n
    Create systemd override:<\/p>\n\n\n\n
    sudo systemctl edit ollama\n<\/code><\/pre>\n\n\n\n
    Paste:<\/p>\n\n\n\n
    [Service]\nEnvironment=\"OLLAMA_HOST=0.0.0.0:11434\"\nEnvironment=\"OLLAMA_CONTEXT_LENGTH=8192\"\nEnvironment=\"OLLAMA_NUM_PARALLEL=1\"\nEnvironment=\"OLLAMA_MAX_LOADED_MODELS=1\"\nEnvironment=\"OLLAMA_KEEP_ALIVE=30m\"\nEnvironment=\"OLLAMA_NO_CLOUD=1\"\n<\/code><\/pre>\n\n\n\n
    Restart:<\/p>\n\n\n\n
    sudo systemctl daemon-reload\nsudo systemctl restart ollama\nsudo systemctl status ollama\n<\/code><\/pre>\n\n\n\n
    Allow LAN access on Linux firewall:<\/p>\n\n\n\n
    sudo ufw allow from 192.168.0.0\/16 to any port 11434 proto tcp\nsudo ufw reload\n<\/code><\/pre>\n\n\n\n
    Do not<\/strong> expose port 11434<\/code> directly to the public internet. Keep it LAN-only or VPN-only.<\/p>\n\n\n\n
    \n\n\n\n
    5. Pull Qwen models on each laptop<\/h1>\n\n\n\n
    On M4 MacBook workers<\/h2>\n\n\n\n
    Try the stronger model first:<\/p>\n\n\n\n
    ollama pull qwen3-coder:30b\n<\/code><\/pre>\n\n\n\n
    Also pull a safer fallback:<\/p>\n\n\n\n
    ollama pull qwen2.5-coder:14b\n<\/code><\/pre>\n\n\n\n
    Test:<\/p>\n\n\n\n
    ollama run qwen3-coder:30b \"Reply only: READY\"\n<\/code><\/pre>\n\n\n\n
    Check memory\/offload:<\/p>\n\n\n\n
    ollama ps\n<\/code><\/pre>\n\n\n\n
    If the Mac becomes slow, hot, or memory pressure goes red, use:<\/p>\n\n\n\n
    ollama stop qwen3-coder:30b\nollama run qwen2.5-coder:14b \"Reply only: READY\"\n<\/code><\/pre>\n\n\n\n
    On Intel i7 64GB worker<\/h2>\n\n\n\n
    Without NVIDIA GPU, Intel will likely be much slower. Start with 14B:<\/p>\n\n\n\n
    ollama pull qwen2.5-coder:14b\nollama run qwen2.5-coder:14b \"Reply only: READY\"\n<\/code><\/pre>\n\n\n\n
    Then test 32B only as a deep\/slow worker:<\/p>\n\n\n\n
    ollama pull qwen2.5-coder:32b\nollama run qwen2.5-coder:32b \"Reply only: READY\"\n<\/code><\/pre>\n\n\n\n
    \n\n\n\n
    6. Test each worker from controller<\/h1>\n\n\n\n
    From your controller laptop:<\/p>\n\n\n\n
    curl http:\/\/m4-1.local:11434\/api\/tags\ncurl http:\/\/m4-2.local:11434\/api\/tags\ncurl http:\/\/intel-1.local:11434\/api\/tags\n<\/code><\/pre>\n\n\n\n
    Test OpenAI-compatible chat endpoint. Ollama supports OpenAI-style \/v1\/chat\/completions<\/code>, and the API key is required by clients but ignored by Ollama. (docs.ollama.com<\/a>)<\/p>\n\n\n\n
    curl http:\/\/m4-1.local:11434\/v1\/chat\/completions \\\n  -H \"Content-Type: application\/json\" \\\n  -d '{\n    \"model\": \"qwen2.5-coder:14b\",\n    \"messages\": [\n      {\"role\": \"user\", \"content\": \"Write a Python hello world in one line.\"}\n    ]\n  }'\n<\/code><\/pre>\n\n\n\n
    For Qwen3-Coder:<\/p>\n\n\n\n
    curl http:\/\/m4-1.local:11434\/v1\/chat\/completions \\\n  -H \"Content-Type: application\/json\" \\\n  -d '{\n    \"model\": \"qwen3-coder:30b\",\n    \"messages\": [\n      {\"role\": \"user\", \"content\": \"Explain what a Kubernetes Deployment does in 3 lines.\"}\n    ]\n  }'\n<\/code><\/pre>\n\n\n\n
    \n\n\n\n
    7. Install LiteLLM router on controller<\/h1>\n\n\n\n
    LiteLLM is useful here because it gives you one OpenAI-compatible gateway and can load-balance multiple deployments under the same model group. Its docs describe LiteLLM Proxy as an LLM gateway with OpenAI-style Chat Completions and load balancing across multiple model deployments. (docs.litellm.ai<\/a>)<\/p>\n\n\n\n
    On controller:<\/p>\n\n\n\n
    mkdir -p ~\/qwen-cluster\ncd ~\/qwen-cluster\n<\/code><\/pre>\n\n\n\n
    Install uv<\/code> if you do not have it:<\/p>\n\n\n\n
    curl -LsSf https:\/\/astral.sh\/uv\/install.sh | sh\nsource ~\/.zshrc 2>\/dev\/null || source ~\/.bashrc 2>\/dev\/null || true\n<\/code><\/pre>\n\n\n\n
    Install LiteLLM Proxy:<\/p>\n\n\n\n
    uv tool install 'litellm[proxy]'\n<\/code><\/pre>\n\n\n\n
    Verify:<\/p>\n\n\n\n
    litellm --version\n<\/code><\/pre>\n\n\n\n
    \n\n\n\n
    8. Create LiteLLM cluster config<\/h1>\n\n\n\n
    Create config:<\/p>\n\n\n\n
    cd ~\/qwen-cluster\nnano litellm-config.yaml\n<\/code><\/pre>\n\n\n\n
    Paste this first version:<\/p>\n\n\n\n
    model_list:\n  # Fast cluster: two M4 laptops running same model\n  - model_name: qwen-coder-fast\n    litellm_params:\n      model: ollama_chat\/qwen2.5-coder:14b\n      api_base: http:\/\/m4-1.local:11434\n      rpm: 4\n\n  - model_name: qwen-coder-fast\n    litellm_params:\n      model: ollama_chat\/qwen2.5-coder:14b\n      api_base: http:\/\/m4-2.local:11434\n      rpm: 4\n\n  # Stronger but tighter M4 model\n  - model_name: qwen-coder-30b\n    litellm_params:\n      model: ollama_chat\/qwen3-coder:30b\n      api_base: http:\/\/m4-1.local:11434\n      rpm: 2\n\n  - model_name: qwen-coder-30b\n    litellm_params:\n      model: ollama_chat\/qwen3-coder:30b\n      api_base: http:\/\/m4-2.local:11434\n      rpm: 2\n\n  # Intel deep\/slow worker\n  - model_name: qwen-coder-intel\n    litellm_params:\n      model: ollama_chat\/qwen2.5-coder:14b\n      api_base: http:\/\/intel-1.local:11434\n      rpm: 1\n\nrouter_settings:\n  routing_strategy: simple-shuffle\n  num_retries: 1\n  timeout: 600\n<\/code><\/pre>\n\n\n\n
    Why separate model groups?<\/p>\n\n\n\n
    qwen-coder-fast  = stable daily coding\nqwen-coder-30b   = better coding, heavier memory\nqwen-coder-intel = slow fallback \/ review \/ tests\n<\/code><\/pre>\n\n\n\n
    Do not<\/strong> mix 14b<\/code>, 30b<\/code>, and 32b<\/code> under the same alias at first. It makes performance unpredictable.<\/p>\n\n\n\n
    \n\n\n\n
    9. Start LiteLLM router<\/h1>\n\n\n\n
    cd ~\/qwen-cluster\nexport LITELLM_MASTER_KEY=\"sk-local-qwen-router-change-me\"\nlitellm --config litellm-config.yaml --host 0.0.0.0 --port 4000\n<\/code><\/pre>\n\n\n\n
    In another terminal, test:<\/p>\n\n\n\n
    curl http:\/\/localhost:4000\/chat\/completions \\\n  -H \"Content-Type: application\/json\" \\\n  -H \"Authorization: Bearer sk-local-qwen-router-change-me\" \\\n  -d '{\n    \"model\": \"qwen-coder-fast\",\n    \"messages\": [\n      {\"role\": \"user\", \"content\": \"Write a Bash command to list files larger than 100MB.\"}\n    ]\n  }'\n<\/code><\/pre>\n\n\n\n
    Test the 30B group:<\/p>\n\n\n\n
    curl http:\/\/localhost:4000\/chat\/completions \\\n  -H \"Content-Type: application\/json\" \\\n  -H \"Authorization: Bearer sk-local-qwen-router-change-me\" \\\n  -d '{\n    \"model\": \"qwen-coder-30b\",\n    \"messages\": [\n      {\"role\": \"user\", \"content\": \"Review this Terraform snippet conceptually: resource aws_s3_bucket test {}\"}\n    ]\n  }'\n<\/code><\/pre>\n\n\n\n
    Test from another laptop:<\/p>\n\n\n\n
    curl http:\/\/controller.local:4000\/chat\/completions \\\n  -H \"Content-Type: application\/json\" \\\n  -H \"Authorization: Bearer sk-local-qwen-router-change-me\" \\\n  -d '{\n    \"model\": \"qwen-coder-fast\",\n    \"messages\": [\n      {\"role\": \"user\", \"content\": \"Reply only: CLUSTER_OK\"}\n    ]\n  }'\n<\/code><\/pre>\n\n\n\n
    \n\n\n\n
    10. Install Qwen Code on controller<\/h1>\n\n\n\n
    Qwen Code is now a proper terminal AI coding agent. Its repo says it supports OpenAI-compatible providers and local models such as Ollama\/vLLM, and it has interactive, headless, IDE, daemon, SDK, and agent-team style usage. (GitHub<\/a>)<\/p>\n\n\n\n
    Install:<\/p>\n\n\n\n
    brew install qwen-code\n<\/code><\/pre>\n\n\n\n
    Or via npm:<\/p>\n\n\n\n
    npm install -g @qwen-code\/qwen-code@latest\n<\/code><\/pre>\n\n\n\n
    Verify:<\/p>\n\n\n\n
    qwen --version\n<\/code><\/pre>\n\n\n\n
    \n\n\n\n
    11. Configure Qwen Code to use your local cluster<\/h1>\n\n\n\n
    Qwen Code supports ~\/.qwen\/settings.json<\/code> with modelProviders<\/code>, security.auth.selectedType<\/code>, and model.name<\/code>. Its docs also show local\/self-hosted models via OpenAI-compatible API by setting baseUrl<\/code>. (Qwen<\/a>)<\/p>\n\n\n\n
    Create config:<\/p>\n\n\n\n
    mkdir -p ~\/.qwen\nnano ~\/.qwen\/settings.json\n<\/code><\/pre>\n\n\n\n
    Paste:<\/p>\n\n\n\n
    {\n  \"modelProviders\": {\n    \"openai\": [\n      {\n        \"id\": \"qwen-coder-fast\",\n        \"name\": \"Local Qwen Coder Fast Cluster\",\n        \"baseUrl\": \"http:\/\/localhost:4000\",\n        \"envKey\": \"LOCAL_QWEN_ROUTER_KEY\",\n        \"generationConfig\": {\n          \"timeout\": 600000,\n          \"maxRetries\": 1,\n          \"contextWindowSize\": 8192,\n          \"samplingParams\": {\n            \"temperature\": 0.2,\n            \"top_p\": 0.9,\n            \"max_tokens\": 4096\n          }\n        }\n      },\n      {\n        \"id\": \"qwen-coder-30b\",\n        \"name\": \"Local Qwen3 Coder 30B Cluster\",\n        \"baseUrl\": \"http:\/\/localhost:4000\",\n        \"envKey\": \"LOCAL_QWEN_ROUTER_KEY\",\n        \"generationConfig\": {\n          \"timeout\": 900000,\n          \"maxRetries\": 1,\n          \"contextWindowSize\": 8192,\n          \"samplingParams\": {\n            \"temperature\": 0.2,\n            \"top_p\": 0.9,\n            \"max_tokens\": 4096\n          }\n        }\n      },\n      {\n        \"id\": \"qwen-coder-intel\",\n        \"name\": \"Local Qwen Intel Worker\",\n        \"baseUrl\": \"http:\/\/localhost:4000\",\n        \"envKey\": \"LOCAL_QWEN_ROUTER_KEY\",\n        \"generationConfig\": {\n          \"timeout\": 1200000,\n          \"maxRetries\": 0,\n          \"contextWindowSize\": 8192,\n          \"samplingParams\": {\n            \"temperature\": 0.2,\n            \"top_p\": 0.9,\n            \"max_tokens\": 4096\n          }\n        }\n      }\n    ]\n  },\n  \"security\": {\n    \"auth\": {\n      \"selectedType\": \"openai\"\n    }\n  },\n  \"model\": {\n    \"name\": \"qwen-coder-fast\"\n  }\n}\n<\/code><\/pre>\n\n\n\n
    Set key:<\/p>\n\n\n\n
    echo 'export LOCAL_QWEN_ROUTER_KEY=\"sk-local-qwen-router-change-me\"' >> ~\/.zshrc\nsource ~\/.zshrc\n<\/code><\/pre>\n\n\n\n
    Launch:<\/p>\n\n\n\n
    qwen\n<\/code><\/pre>\n\n\n\n
    Inside Qwen Code:<\/p>\n\n\n\n
    \/model\n<\/code><\/pre>\n\n\n\n
    Select:<\/p>\n\n\n\n
    Local Qwen Coder Fast Cluster\n<\/code><\/pre>\n\n\n\n
    Test prompt:<\/p>\n\n\n\n
    Analyze this repository and tell me the tech stack, entry points, and test commands.\n<\/code><\/pre>\n\n\n\n
    \n\n\n\n
    12. Use it for real coding<\/h1>\n\n\n\n
    Create a test repo:<\/p>\n\n\n\n
    mkdir -p ~\/ai-lab\ncd ~\/ai-lab\ngit clone <YOUR_REPO_URL> app\ncd app\n<\/code><\/pre>\n\n\n\n
    Run Qwen Code:<\/p>\n\n\n\n
    qwen\n<\/code><\/pre>\n\n\n\n
    Good first prompt:<\/p>\n\n\n\n
    Do not modify files yet. First inspect the repo and tell me:\n1. tech stack\n2. build command\n3. test command\n4. risky files\n5. suggested first small task\n<\/code><\/pre>\n\n\n\n
    Then:<\/p>\n\n\n\n
    Create a plan to add unit tests for the authentication module. Do not execute commands without asking.\n<\/code><\/pre>\n\n\n\n
    Then:<\/p>\n\n\n\n
    Implement the first small test only. Keep changes minimal.\n<\/code><\/pre>\n\n\n\n
    Review:<\/p>\n\n\n\n
    git diff\ngit status\n<\/code><\/pre>\n\n\n\n
    Commit only after review:<\/p>\n\n\n\n
    git add .\ngit commit -m \"test: add auth unit test\"\n<\/code><\/pre>\n\n\n\n
    \n\n\n\n
    13. Turn it into a real multi-agent coding cluster<\/h1>\n\n\n\n
    This is where your 3\u20134 laptops become useful.<\/p>\n\n\n\n
    Use Git worktrees so each AI agent works separately.<\/p>\n\n\n\n
    cd ~\/ai-lab\/app\n\ngit worktree add ..\/app-agent-1 -b ai\/agent-1-auth-tests\ngit worktree add ..\/app-agent-2 -b ai\/agent-2-docs\ngit worktree add ..\/app-agent-3 -b ai\/agent-3-refactor\n<\/code><\/pre>\n\n\n\n
    Open 3 terminals.<\/p>\n\n\n\n
    Terminal 1<\/h2>\n\n\n\n
    cd ~\/ai-lab\/app-agent-1\nqwen -p \"You are agent 1. Add focused unit tests for the authentication module. Keep changes minimal. Run tests if possible. Do not touch unrelated files.\"\n<\/code><\/pre>\n\n\n\n
    Terminal 2<\/h2>\n\n\n\n
    cd ~\/ai-lab\/app-agent-2\nqwen -p \"You are agent 2. Improve README developer setup instructions based on this repository. Do not change source code.\"\n<\/code><\/pre>\n\n\n\n
    Terminal 3<\/h2>\n\n\n\n
    cd ~\/ai-lab\/app-agent-3\nqwen -p \"You are agent 3. Find one small refactor opportunity in the API layer. Make a minimal safe change and explain the risk.\"\n<\/code><\/pre>\n\n\n\n
    Qwen Code\u2019s docs list headless mode as qwen -p \"...\"<\/code>, useful for scripts, CI\/CD, and batch processing. (GitHub<\/a>)<\/p>\n\n\n\n
    Now your agents are parallel:<\/p>\n\n\n\n
    Agent 1 \u2192 test branch\nAgent 2 \u2192 docs branch\nAgent 3 \u2192 refactor branch\n<\/code><\/pre>\n\n\n\n
    Check results:<\/p>\n\n\n\n
    cd ~\/ai-lab\/app-agent-1\ngit diff\n\ncd ~\/ai-lab\/app-agent-2\ngit diff\n\ncd ~\/ai-lab\/app-agent-3\ngit diff\n<\/code><\/pre>\n\n\n\n
    Merge manually after review.<\/p>\n\n\n\n
    \n\n\n\n
    14. Optional: create a simple task runner<\/h1>\n\n\n\n
    Create file:<\/p>\n\n\n\n
    cd ~\/qwen-cluster\nnano run-agent-task.sh\n<\/code><\/pre>\n\n\n\n
    Paste:<\/p>\n\n\n\n
    #!\/usr\/bin\/env bash\nset -euo pipefail\n\nREPO=\"$1\"\nBRANCH=\"$2\"\nTASK=\"$3\"\n\nBASE_DIR=\"$HOME\/ai-lab\"\nWORKTREE=\"$BASE_DIR\/$BRANCH\"\n\nmkdir -p \"$BASE_DIR\"\n\ncd \"$REPO\"\n\ngit fetch --all --prune\ngit worktree add \"$WORKTREE\" -b \"$BRANCH\" || true\n\ncd \"$WORKTREE\"\n\nqwen -p \"$TASK\"\n\necho\necho \"===== STATUS =====\"\ngit status --short\n\necho\necho \"===== DIFF =====\"\ngit diff --stat\n<\/code><\/pre>\n\n\n\n
    Make executable:<\/p>\n\n\n\n
    chmod +x run-agent-task.sh\n<\/code><\/pre>\n\n\n\n
    Run:<\/p>\n\n\n\n
    ~\/qwen-cluster\/run-agent-task.sh \\\n  ~\/ai-lab\/app \\\n  ai\/add-healthcheck-tests \\\n  \"Add tests for the healthcheck endpoint. Keep the change minimal. Run the relevant test command if you can infer it.\"\n<\/code><\/pre>\n\n\n\n
    Run another:<\/p>\n\n\n\n
    ~\/qwen-cluster\/run-agent-task.sh \\\n  ~\/ai-lab\/app \\\n  ai\/update-deployment-docs \\\n  \"Update deployment documentation. Do not change application source code.\"\n<\/code><\/pre>\n\n\n\n
    This is your basic AI coding farm.<\/p>\n\n\n\n
    \n\n\n\n
    15. Optional: use Aider with the same cluster<\/h1>\n\n\n\n
    Aider can connect to any OpenAI-compatible endpoint, including local endpoints. Its docs show OPENAI_API_BASE<\/code> and OPENAI_API_KEY<\/code> for OpenAI-compatible APIs. (aider.chat<\/a>)<\/p>\n\n\n\n
    Install:<\/p>\n\n\n\n
    python3 -m pip install aider-install\naider-install\n<\/code><\/pre>\n\n\n\n
    Configure:<\/p>\n\n\n\n
    export OPENAI_API_BASE=\"http:\/\/localhost:4000\"\nexport OPENAI_API_KEY=\"sk-local-qwen-router-change-me\"\n<\/code><\/pre>\n\n\n\n
    Run:<\/p>\n\n\n\n
    cd ~\/ai-lab\/app\naider --model openai\/qwen-coder-fast\n<\/code><\/pre>\n\n\n\n
    Or try 30B:<\/p>\n\n\n\n
    aider --model openai\/qwen-coder-30b\n<\/code><\/pre>\n\n\n\n
    \n\n\n\n
    16. Performance tuning<\/h1>\n\n\n\n
    For 24GB M4 MacBooks<\/h2>\n\n\n\n
    Start safe:<\/p>\n\n\n\n
    launchctl setenv OLLAMA_CONTEXT_LENGTH \"8192\"\nlaunchctl setenv OLLAMA_NUM_PARALLEL \"1\"\nlaunchctl setenv OLLAMA_MAX_LOADED_MODELS \"1\"\n<\/code><\/pre>\n\n\n\n
    For qwen2.5-coder:14b<\/code>, you can try:<\/p>\n\n\n\n
    launchctl setenv OLLAMA_CONTEXT_LENGTH \"16384\"\n<\/code><\/pre>\n\n\n\n
    For qwen3-coder:30b<\/code>, stay conservative:<\/p>\n\n\n\n
    launchctl setenv OLLAMA_CONTEXT_LENGTH \"8192\"\n<\/code><\/pre>\n\n\n\n
    Then restart Ollama:<\/p>\n\n\n\n
    osascript -e 'quit app \"Ollama\"' || true\nopen -a Ollama\n<\/code><\/pre>\n\n\n\n
    Check:<\/p>\n\n\n\n
    ollama ps\n<\/code><\/pre>\n\n\n\n
    For Linux Intel<\/h2>\n\n\n\n
    If CPU-only, do not expect magic. Use it for:<\/p>\n\n\n\n
    slow review\nsmall model\ntest runner\nDocker\ndatabase\nGit worktree storage\n<\/code><\/pre>\n\n\n\n
    Use:<\/p>\n\n\n\n
    sudo systemctl edit ollama\n<\/code><\/pre>\n\n\n\n
    Keep:<\/p>\n\n\n\n
    [Service]\nEnvironment=\"OLLAMA_CONTEXT_LENGTH=8192\"\nEnvironment=\"OLLAMA_NUM_PARALLEL=1\"\nEnvironment=\"OLLAMA_MAX_LOADED_MODELS=1\"\n<\/code><\/pre>\n\n\n\n
    Restart:<\/p>\n\n\n\n
    sudo systemctl daemon-reload\nsudo systemctl restart ollama\n<\/code><\/pre>\n\n\n\n
    \n\n\n\n
    17. Security checklist<\/h1>\n\n\n\n
    This part matters. Local coding agents can read files, run shell commands, and modify repos.<\/p>\n\n\n\n
    Do this:<\/p>\n\n\n\n
    Use separate repo clones\nUse Git worktrees\nUse branch per task\nDo not mount production secrets\nDo not expose Ollama\/LiteLLM to internet\nUse LAN\/VPN only\nUse LiteLLM master key\nReview git diff before commit\nNever let agent push directly to main\nNever give AWS\/GCP production credentials to agent shell\n<\/code><\/pre>\n\n\n\n
    Also enable local-only Ollama mode:<\/p>\n\n\n\n
    launchctl setenv OLLAMA_NO_CLOUD \"1\"\n<\/code><\/pre>\n\n\n\n
    or on Linux:<\/p>\n\n\n\n
    Environment=\"OLLAMA_NO_CLOUD=1\"\n<\/code><\/pre>\n\n\n\n
    Ollama says local prompts\/data are not sent to Ollama when running locally, and cloud features can be disabled with OLLAMA_NO_CLOUD=1<\/code> or server config. (docs.ollama.com<\/a>)<\/p>\n\n\n\n
    \n\n\n\n
    18. Troubleshooting<\/h1>\n\n\n\n
    Problem: curl http:\/\/m4-1.local:11434\/api\/tags<\/code> fails<\/h2>\n\n\n\n
    Check Ollama is listening:<\/p>\n\n\n\n
    lsof -i :11434\n<\/code><\/pre>\n\n\n\n
    On macOS:<\/p>\n\n\n\n
     listening:\n\n```bash\nlaunchctl getenv OLLAMA_HOST\n<\/code><\/pre>\n\n\n\n
    Restart:<\/p>\n\n\n\n
    osascript -e 'quit app \"Ollama\"' || true\nopen -a Ollama\n<\/code><\/pre>\n\n\n\n
    On Linux:<\/p>\n\n\n\n
    sudo systemctl status ollama\njournalctl -u ollama -n 100 --no-pager\n<\/code><\/pre>\n\n\n\n
    Problem: model too slow<\/h2>\n\n\n\n
    Use smaller model:<\/p>\n\n\n\n
    ollama pull qwen2.5-coder:7b\nollama run qwen2.5-coder:7b \"Reply READY\"\n<\/code><\/pre>\n\n\n\n
    Then update LiteLLM config:<\/p>\n\n\n\n
    model: ollama_chat\/qwen2.5-coder:7b\n<\/code><\/pre>\n\n\n\n
    Restart LiteLLM.<\/p>\n\n\n\n
    Problem: Mac memory pressure<\/h2>\n\n\n\n
    Stop heavy model:<\/p>\n\n\n\n
    ollama stop qwen3-coder:30b\n<\/code><\/pre>\n\n\n\n
    Use 14B:<\/p>\n\n\n\n
    ollama run qwen2.5-coder:14b\n<\/code><\/pre>\n\n\n\n
    Problem: LiteLLM not routing<\/h2>\n\n\n\n
    Run debug:<\/p>\n\n\n\n
    export LITELLM_LOG=DEBUG\nlitellm --config litellm-config.yaml --host 0.0.0.0 --port 4000 --detailed_debug\n<\/code><\/pre>\n\n\n\n
    LiteLLM docs show --detailed_debug<\/code> for debug(docs.litellm.ai<\/a>)cite\ue202turn843007view1\ue201<\/p>\n\n\n\n
    Problem: Qwen Code opens wrong auth flow<\/h2>\n\n\n\n
    Force OpenAI-compatible auth in:<\/p>\n\n\n\n
    nano ~\/.qwen\/settings.json\n<\/code><\/pre>\n\n\n\n
    Ensure:<\/p>\n\n\n\n
    \"security\": {\n  \"auth\": {\n    \"selectedType\": \"openai\"\n  }\n}\n<\/code><\/pre>\n\n\n\n
    Qwen Code docs say security.auth.selectedType<\/code> controls which protocol is used at startup, and model.name<\/code> must match one config(Qwen<\/a>)cite\ue202turn383593view1\ue201<\/p>\n\n\n\n
    \n\n\n\n
    Final recommended setup for your exact hardware<\/h1>\n\n\n\n
    Use this first:<\/p>\n\n\n\n
    M4 MacBook #1:\n  Ollama + qwen2.5-coder:14b\n  Later try qwen3-coder:30b\n\nM4 MacBook #2:\n  Ollama + qwen2.5-coder:14b\n  Later try qwen3-coder:30b\n\nIntel i7 64GB:\n  Ollama + qwen2.5-coder:14b\n  LiteLLM router\n  Git worktrees\n  Docker\/test runner\n\nController:\n  Prefer Intel if always powered on\n  Prefer M4 if you want smoother interactive Qwen Code\n<\/code><\/pre>\n\n\n\n
    Best stable model group:<\/p>\n\n\n\n
    qwen-coder-fast = two M4 laptops running qwen2.5-coder:14b\n<\/code><\/pre>\n\n\n\n
    Best experimental model group:<\/p>\n\n\n\n
    qwen-coder-30b = two M4 laptops running qwen3-coder:30b with 8K context\n<\/code><\/pre>\n\n\n\n
    The real productivity boost will come from:<\/p>\n\n\n\n
    1 repo\n3 worktrees\n3 Qwen Code headless agents\n1 LiteLLM router\n2\u20133 local Qwen model workers\nhuman review before merge\n<\/code><\/pre>\n\n\n\n
    That is the practical self-hostedQwen coding cluster.<\/p>\n","protected":false},"excerpt":{"rendered":"
    Below is a complete practical tutorial. Self-Hosted Qwen Coding Cluster Using 3\u20134 Laptops 0. What we are building You will […]<\/p>\n","protected":false},"author":4,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-3719","post","type-post","status-publish","format-standard","hentry","category-uncategorized"],"_links":{"self":[{"href":"https:\/\/aiopsschool.com\/blog\/wp-json\/wp\/v2\/posts\/3719","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/aiopsschool.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/aiopsschool.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/aiopsschool.com\/blog\/wp-json\/wp\/v2\/users\/4"}],"replies":[{"embeddable":true,"href":"https:\/\/aiopsschool.com\/blog\/wp-json\/wp\/v2\/comments?post=3719"}],"version-history":[{"count":1,"href":"https:\/\/aiopsschool.com\/blog\/wp-json\/wp\/v2\/posts\/3719\/revisions"}],"predecessor-version":[{"id":3720,"href":"https:\/\/aiopsschool.com\/blog\/wp-json\/wp\/v2\/posts\/3719\/revisions\/3720"}],"wp:attachment":[{"href":"https:\/\/aiopsschool.com\/blog\/wp-json\/wp\/v2\/media?parent=3719"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/aiopsschool.com\/blog\/wp-json\/wp\/v2\/categories?post=3719"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/aiopsschool.com\/blog\/wp-json\/wp\/v2\/tags?post=3719"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}