Qwen 3.6 27B Local LLM’s TitleForge Performance: Replacing Claude Code
Generated: 2026-07-14 · API: Gemini 2.5 Flash · Modes: Summary
Qwen 3.6 27B Local LLM’s TitleForge Performance: Replacing Claude Code
Clip title: Can Qwen 3.6 27B Replace Claude Code? Full Uncut Local Coding Session Author / channel: CloudYeti | AI Engineering URL: https://www.youtube.com/watch?v=6NhLP_YGZVw
Summary
This video demonstrates a real-time coding challenge involving the Qwen 3.6 27B 6-bit quantized model, running entirely locally on a 128GB Mac. The objective was to build a full-stack web application called “TitleForge,” capable of generating YouTube video titles and thumbnails. A key constraint for this project was the exclusive use of Codex CLI for all large language model (LLM) inference and image generation, thereby avoiding external APIs like OpenAI or Claude. This specific requirement was designed to rigorously test the capabilities of local LLMs for complex application development.
Initially, Qwen faced several obstacles, including title generation failures, server crashes, and an tendency to rewrite entire code files rather than applying small, precise edits—a behavior the presenter noted as a drawback of the Pi coding agent used. The video also showcased the significant resource consumption, with RAM usage climbing to 65GB and CPU fan noise indicating high utilization. Despite these challenges, Qwen exhibited remarkable autonomy, debugging and correcting its own errors without any external human intervention, a notable improvement compared to a previous test with DeepSeek which required a hint from GPT.
Ultimately, Qwen successfully built and deployed the “TitleForge” application. The finished product features a clean, minimalist user interface, similar to a prior DeepSeek-generated app, and effectively generates clickable, ranked titles with associated scores, along with high-quality, relevant thumbnails. The demonstration highlights several key takeaways: local LLMs, particularly advanced models like Qwen, are capable of building sophisticated applications with specific local execution requirements. However, this comes with considerable demands on computational resources and can involve iterative debugging. The choice of coding agent and quantization methods (like Q6 and MTP) are crucial for balancing output quality and speed when working with local AI development.
Video Description & Links
Description
Qwen 3.6 27B (6-bit quant, MTP) builds a complete full-stack app on my 128GB M3 Max Mac in one uncut session — about 52 minutes, with zero help from ChatGPT or Claude. Last time I gave DeepSeek the same test, it took two hours and needed a hint from GPT.
The twist: the app itself isn’t allowed to call OpenAI or any cloud API directly. It has to generate its titles and thumbnails through codex exec, the command-line side of my Codex subscription — a setup the model has almost certainly never seen. You’ll watch it fail the first generation, crash the server, debug it, and finish anyway. RAM and CPU stay on screen the whole run, since a bunch of you asked for that last video.
Model: Qwen 3.6 27B, Q6 quant with multi-token prediction (MTP), served via llama.cpp, driven by the Pi coding agent.
CHAPTERS
00:00 What happened with DeepSeek + this test 01:46 The app we’re building (and the twist) 03:12 Starting Pi — Qwen 3.6 27B Q6 MTP 04:07 The codex exec handoff doc 07:09 RAM + CPU on screen (you asked for this) 20:43 Dependencies install — so far so smooth 22:50 Why DeepSeek took 2 hours on this same test 24:00 First generation fails 27:18 Giving up the World Cup quarterfinal for this 44:10 Progress check at 42 minutes 51:15 “It works now” — the moment it clicks 53:44 Generating the thumbnail 54:29 52 minutes, uncut 55:58 The model setup, recapped (Q6 + MTP) 56:54 Work with me
RELATED VIDEOS
DeepSeek, same test, 2 hours (uncut): https://youtu.be/hVpOxnESKVs DeepSeek V4 Flash on a 128GB Mac (DS4): https://youtu.be/Xd6zvPk9X5M Best local LLM for a 128GB Mac: https://youtu.be/wU8KSIU-wIk Local LLM playlist: https://www.youtube.com/playlist?list=PLQP5dDPLts67psUzW096fJdMNHJB16Sob
Fina code repo on GITHUB: https://github.com/ravsau/qwen-3-6-full-stack
WORK WITH CLOUDYETI
If your team or company wants help adopting AI, using it well, or cutting AI and cloud costs, book a free 15-minute call: https://cloudyeti.io/chat
If you want hands-on help with your own setup, book a paid 1:1: https://cloudyeti.io/meet
Tags
128gb mac, apple silicon llm, codex cli, codex exec, full stack app with ai, llama.cpp, local ai coding, local coding agent, local llm, m3 max, mtp, pi coding agent, qwen 3.6, qwen 3.6 27b, uncut coding session
URLs
- https://youtu.be/hVpOxnESKVs
- https://youtu.be/Xd6zvPk9X5M
- https://youtu.be/wU8KSIU-wIk
- https://www.youtube.com/playlist?list=PLQP5dDPLts67psUzW096fJdMNHJB16Sob
- https://github.com/ravsau/qwen-3-6-full-stack
- https://cloudyeti.io/chat
- https://cloudyeti.io/meet
- https://x.com/ravsau