The latest articles on DEV Community by nerudek_vibecoder (@nerua1). https://dev.to/nerua1 https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F3911043%2F7ebcb4a0-de99-4a77-ab63-61d3cdfd6ebf.jpeg https://dev.to/nerua1 en nerudek_vibecoder Sun, 03 May 2026 23:43:51 +0000 https://dev.to/nerua1/how-i-built-a-multi-model-ai-council-that-runs-on-a-mac-mini-49b8 https://dev.to/nerua1/how-i-built-a-multi-model-ai-council-that-runs-on-a-mac-mini-49b8 <h1> How I Built a Multi-Model AI Council That Runs on a Mac Mini </h1> <p>I run 4 AI agents (Claude Code, OpenClaw, Hermes/DeepSeek, LM Studio) on a single Mac Mini M4 with 32GB RAM. They share memory through Obsidian + ChromaDB, communicate via ACP bridge, and delegate tasks using a tiered hierarchy. Here's what actually works and what breaks.</p> <h2> The Stack </h2> <ul> <li> <strong>Orchestrator</strong>: DeepSeek V4 Pro (API) — plans, delegates, communicates</li> <li> <strong>Architect</strong>: Claude Code — design decisions, code review</li> <li> <strong>Workers</strong>: LM Studio with qwen3.5-27b — heavy coding, research</li> <li> <strong>Memory</strong>: Obsidian vault + MemPalace (ChromaDB) — shared across all agents</li> <li> <strong>Communication</strong>: acpx bridge — <code>npx acpx claude exec "message"</code> </li> </ul> <h2> The Hard Part: RAM At 32GB </h2> <p>32GB minus ~7GB for macOS = 25GB usable. One large model (~22GB) leaves almost nothing for subagents. The solution: cloud orchestrator (DeepSeek) + local workers. The orchestrator costs ~$3/month in API tokens. Workers cost electricity (which runs anyway).</p> <h2> Arena Council: Multi-Model Voting </h2> <p>When one model isn't enough, I run 3-5 local models in parallel and let them vote. Each model gets the same query, responses are scored, and the best one wins. It's like having a panel of experts instead of one.<br> </p> <div class="highlight js-code-highlight"> <pre class="highlight python"><code><span class="c1"># 5 models, 1 answer </span><span class="n">council</span> <span class="o">=</span> <span class="nc">ArenaCouncil</span><span class="p">(</span><span class="n">models</span><span class="o">=</span><span class="p">[</span><span class="sh">"</span><span class="s">qwen-27b</span><span class="sh">"</span><span class="p">,</span> <span class="sh">"</span><span class="s">qwen-35b</span><span class="sh">"</span><span class="p">,</span> <span class="sh">"</span><span class="s">mistral-7b</span><span class="sh">"</span><span class="p">,</span> <span class="sh">"</span><span class="s">llama-8b</span><span class="sh">"</span><span class="p">,</span> <span class="sh">"</span><span class="s">qwen-9b</span><span class="sh">"</span><span class="p">])</span> <span class="n">result</span> <span class="o">=</span> <span class="n">council</span><span class="p">.</span><span class="nf">ask</span><span class="p">(</span><span class="sh">"</span><span class="s">Should I refactor this monorepo?</span><span class="sh">"</span><span class="p">)</span> <span class="c1"># → 3/5 say yes → majority wins </span></code></pre> </div> <p>All local, all free (beyond electricity).</p> <h2> What Broke (And How I Fixed It) </h2> <ol> <li> <strong>Sandbox mode</strong> — killed symlinks to external drive. Disabled it.</li> <li> <strong>LM Studio kernel panics</strong> — loading 2 large models simultaneously on M4 GPU. Now max 1 large model.</li> <li> <strong>Doctor --repair</strong> — overwrote config files. Now backup before any repair.</li> <li> <strong>Lobotomy</strong> — agent overwrote its own personality file. Added file-level write protection.</li> </ol> <p>Full story + code: <a href="https://github.com/nerua1" rel="noopener noreferrer">github.com/nerua1</a></p> <p><em>If this saved you time: <a href="https://www.paypal.me/nerudek" rel="noopener noreferrer">☕ PayPal.me/nerudek</a></em></p> ai opensource llm tutorial