Apple's OFFICIAL Local LLM AI Coding in Xcode, MLX + Qwen 3.6 & Gemma 4 on 32GB M5 Macbook Pro
Apple just made it OFFICIAL — on their website and their official Apple Developer YouTube channel — that you can now do local LLM vibe coding directly inside Xcode on your Mac, powered by MLX and Apple Silicon. That means AI-assisted coding, agentic tool calling, and full app generation with NO cloud, NO API keys, NO subscription, and NO data ever leaving your local machine. This is a huge moment for on-device AI, private AI coding, and offline development on macOS.
But here's the honest catch: Apple demoed all of this on a maxed-out M5 MacBook Pro Max with 128GB of unified memory. I'm running an M5 Pro MacBook Pro with only 32GB. So the real question every normal developer is asking — can a regular, non-maxed-out Mac actually run local AI coding in Xcode without crashing? That's exactly what we test in this video, unfiltered and hands-on, the way this channel always does it: only stuff you can genuinely try yourself.
In this local LLM vs Xcode test, I connect LM Studio to Xcode's built-in AI agent, load MLX-optimized models, and actually build and debug a working SwiftUI iOS / iPad To-Do app — then push it further with an Apple PencilKit drawing feature — using nothing but a local open-source model running on my Mac.
LOCAL AI MODELS TESTED IN XCODE:
Qwen 3.6 35B A3B — 4-bit quantized MLX model, Mixture-of-Experts (MoE), with vision, tool calling and reasoning
Gemma 4 26B A4B QAT — the exact model Apple used in their official demo
Gemma 4 12B QAT — smaller quantization-aware model
Xcode's built-in free ChatGPT — used as a fallback comparison against local models
WHAT YOU'LL LEARN:
How to set up Xcode with a local LLM step by step
How to add LM Studio (or Ollama) as a local chat provider inside Xcode's agent settings
What a QAT (Quantization-Aware Training) model is and why it keeps quality close to BF16
What MLX is, how the MLX agentic loop calling works, and what the latest MLX update changed for Apple Silicon
Where the KV cache actually lives — SSD swap vs unified memory — and why that's what stops a 32GB Mac from crashing
Real-world memory pressure, unified memory usage, failures, retries, pre-warming the model, and context window tuning (20K vs 128K)
How Apple's orchestration layer feeds project context to the model and drives Xcode's tools
Building an iPad drawing app with PencilKit, generated entirely by a local AI model
Whether local LLM coding is actually reliable right now, or still early
THE HARDWARE:
M5 Pro MacBook Pro, 32GB unified memory, Apple Silicon — tested against Apple's 128GB M5 MacBook Pro Max demo. If you want to know whether this works on a 16GB M1 / M2 / M3 Mac, or on a Windows PC with an NVIDIA RTX graphics card, comment below and I'll cover it.
This is honest, no-hype local AI testing: local LLM vibe coding, on-device coding on Mac, private offline app development, and a real look at whether Apple Intelligence + MLX + open models can replace cloud AI coding tools for indie developers and Swift developers.
Timestamps
00:00 – Huge news: Apple officially backs local LLM vibe coding
00:39 – Xcode's built-in chat / agent window
01:14 – Adding LM Studio as a local chat provider
01:45 – The local models available (Gemma 4 & Qwen 3.6 QAT / MoE)
02:09 – The real test: 32GB M5 Pro vs Apple's 128GB M5 Max
02:46 – First prompt: build a To-Do app with Qwen 3.6 35B (MLX)
03:29 – MLX model breakdown (4-bit, MoE, vision, tool calling, reasoning)
05:12 – MLX update + the KV cache / unified memory / SSD deep dive
07:03 – First attempt fails (LM Studio fetch error)
07:49 – Asking Claude for help + pre-warming the model
08:36 – Ejecting, reloading, and bumping the context window to 128K
10:19 – It works! The To-Do app is generated and tested
11:37 – Prompting a drawing feature (PencilKit)
12:56 – Hitting the context / memory error again
14:43 – Switching to Gemma 4 QAT (the model Apple used)
14:52 – What is a QAT model, explained
15:25 – Gemma 4 running the agentic loop inside Xcode
16:51 – Apple's orchestration + MLX loop calling doing the heavy lifting
17:53 – The PencilKit drawing app gets built
18:38 – Bug fixing attempts (6 tries)
19:37 – Falling back to Xcode's built-in free ChatGPT
19:57 – Final iPad app demo
20:30 – Verdict: can you rely on local LLM vibe coding on Mac?
21:22 – Outro
If you like honest, real-world local AI testing with no hype — only things you can actually try yourself — please subscribe, like, and share. It genuinely keeps this channel alive.
#LocalLLM #MLX #Xcode #VibeCoding #AppleIntelligence #LocalAI #Qwen3 #Gemma4 #LMStudio #M5MacBook #AppleSilicon #OnDeviceAI #AICoding #PrivateAI #SwiftUI #Ollama #OpenSourceAI #MacBookPro #iOSDev #Coding