This video introduces and provides a hands-on "vibe test" of the newly released Quen 3 coder 30B A3B Instruct model, a 30 billion parameter AI model designed for local coding tasks. The presenter runs the Q8 quantized version on a dual 3090 Ti system using LM Studio. Key highlights include its native 256k context length (extendable to 1M) and support for agentic coding. The testing primarily focuses on generating code for various applications. The model demonstrates impressive capabilities in producing functional and aesthetically pleasing web interfaces using HTML, CSS, and JavaScript, including a browser-based OS simulation and a highly customized "Steve's PC Repair" website that humorously incorporates "Mac hatred." However, its performance in generating Python-based 2D and 3D games is less successful, showing functional limitations and errors. Overall, the video concludes that for a locally runnable, quantized model, its web development and UI design capabilities are highly impressive, making it a promising and accessible tool for many users despite some shortcomings in complex game generation.
Summarizing the Quen 3 Coder 30B A3B Instruct Model
0:00 Introduction of Quen 3 Coder: The video introduces the new 30 billion parameter A3B Instruct version of the Quen 3 coder family, highlighting its potential for local execution on less expensive hardware.
0:07 Model Highlights: This model is designed for agentic coding tasks, boasts a native 256k context length (extendable up to 1 million tokens), and operates in a "non-thinking" mode for immediate output.
0:35 Local Setup: The presenter runs the model locally using a Q8 quantization (32.5 GB download) on a dual 3090 Ti system, loaded in LM Studio with a 32,768 context length.
3:47 Test 1: Browser-Based OS: The model is tasked with generating a browser-based operating system using JavaScript, CSS, and HTML. It produces a functional UI with a working clock, a start menu with applications (calculator, file explorer, settings, browser), and aesthetically pleasing elements, though the calculator is non-functional and window focus isn't perfect.
8:43 Test 2: Python Slot Machine: The model generates a Python-based slot machine game. It functions partially but struggles with displaying emojis (showing Unicode references instead) and exhibits a "glitching" infinite spin, prompting the presenter to troubleshoot by requesting simple text characters instead.
10:39 Test 3: Arithmetic: A simple arithmetic problem is posed to the model (calculating subscribers if all non-subscribers subscribed), which it answers correctly.
11:13 Test 4: Steve's PC Repair Website: The model generates an HTML website for a fictional "Steve's PC Repair." The initial version is well-designed but has image loading issues.
13:22 Test 4.1: Steve's PC Repair Website - Mac Hatred: The model is instructed to modify the website to reflect "Steve's hatred of Macs." It impressively re-themes the site with aggressive red colors, "No Macs Allowed" banners, and humorous, stern disclaimers about not servicing Apple products, even integrating these into the footer.
16:44 Test 5: Johnny's Bitcoin Duplicator Website: The model is asked to create a convincing, flashy "Bitcoin duplicator" website with a private key input. The model generates a visually appealing, casino-themed site with concerningly high "success rates."
18:15 Test 5.1: Bitcoin Duplicator Website - Disclaimer Removal: The model successfully removes all disclaimers from the Bitcoin duplicator site upon request.
18:44 Test 6: 3D Python FPS Game: The model attempts to generate a simple 3D first-person shooter using Python, OpenGL, and Pygame. While it quickly generates code, the game results in an error and shows only partial, rudimentary 3D functionality with clipping issues.
20:44 Overall Impressions: The presenter concludes that the model excels in UI design and web development (HTML, CSS, JavaScript), producing functional and aesthetically pleasing results, even with humorous, specific requests. However, its Python game generation capabilities are less impressive.
22:04 Key Takeaway: The model's ability to run locally on consumer hardware with such strong web development capabilities makes it a highly promising and accessible tool.