MacBook Pro M5 + Qwen 3.6 35B Test | LOCAL AI Vibe Code omlx with vs code

Can a fully local AI model running on a MacBook Pro M5 (32GB) actually vibe code a production-level app — no cloud, no API keys, nothing leaving your machine? I put it to the test. In this video I run Qwen 3.6 35B A3B (4-bit quantized, ~19GB, the new MoE coding model from Alibaba) through an MLX local server, wire it into VS Code with Continue.dev as the chat/agent, and try to build a real app — a platform where AI vibe-coding developers can showcase work and land clients — purely as a "newbie" vibe coder who doesn't touch the code. I also get into why I ditched Ollama and LM Studio for MLX, the context-window crashes that kept killing the model, real tokens/sec numbers, the M5's Low Power Mode situation, and the honest cost math: a ~$2,400 MacBook for local LLMs vs $20/month cloud tools like Claude Code (Opus 4.8) and GitHub Copilot. Does local win in 2026? Stick around for the final verdict. 🛠️ Gear & tools used: • MacBook Pro M5 — 32GB unified memory, 1TB SSD • Model: Qwen 3.6 35B A3B, 4-bit quantized (MLX) • MLX local server (OMLX) • VS Code + Continue.dev (local agent) • Compared against: GitHub Copilot, Claude Code (Opus 4.8 / Sonnet 4.6), LM Studio, Ollama Chapters: 0:00 Intro – can a local LLM build a full app? 0:35 The rig: MacBook Pro M5 (32GB / 1TB SSD) 0:46 Why I dropped Ollama & LM Studio for MLX 2:34 The model: Qwen 3.6 35B A3B (4-bit, 19GB) 3:09 Planning the app – a platform for vibe coders 4:33 Memory use & tokens/sec while it thinks 5:40 Why this is my last time with GitHub Copilot 6:39 Going 100% local: Continue.dev + MLX server 8:36 First errors & the context-window problem 10:29 It starts building the app 12:43 The honest math: M5 (~$2,400) vs $20/mo cloud 16:00 Real token speeds with the MLX dashboard on 22:50 M5's Low Power Mode + speed after closing the dashboard 25:05 The app finally runs (and keeps crashing) 27:14 Final verdict: is local vibe coding worth it? 29:02 What I'm testing next + subscribe New channel, zero filter — I do these experiments live and tell you straight what works. Got a model you want me to test (Gemma 4, Qwen Coder 30B, GPT-OSS 20B)? Drop it in the comments and subscribe so you don't miss it. #localllm #vibecoding #macbookprom5 #qwen3 #aicoding