This video provides a hands-on review and "vibe test" of the newly released Quen 3 Coder 30B A3B Instruct, a 30 billion-parameter large language model designed for local execution on consumer-grade hardware. The presenter evaluates its coding capabilities by tasking it with various projects, including generating a browser-based operating system, a Python slot machine, and a website with a specific, humorous persona. The model demonstrates impressive proficiency in web development (HTML/CSS/JavaScript), producing aesthetically pleasing and largely functional interfaces, and effectively adopting a specified tone. However, its performance in Python game generation was less reliable, with functional issues observed. The video highlights the model's accessibility for local deployment and its strong potential for UI/UX design tasks.
Exploring the Quen 3 Coder 30B A3B Instruct: A Hands-On Review
0:00 New Model for Local Use: The video introduces the Quen 3 Coder 30B A3B Instruct, a new 30 billion-parameter coding model designed to run effectively on local machines, making it accessible to users without high-end hardware.
1:14 Key Model Features: Highlights include strong performance in agentic coding, a native context length of 256k (extendable to 1 million tokens), support for various agentic coding frameworks, and its ability to provide immediate answers without "thinking blocks."
2:26 Testing Environment: The presenter uses a Q8 quantization (32.5 GB download size) of the model on a dual 3090 Ti system with LM Studio, utilizing around 38-39 GB of VRAM and a 32,768 context length.
3:45 Browser OS Generation Test: The model successfully generated a browser-based operating system with functional elements like a clock, file explorer, settings app (with theme and resolution changes), and draggable/resizable windows. While visually appealing and mostly functional, the calculator application was not working.
8:26 Python Slot Machine Game Test: The attempt to create a Python slot machine game was less successful. It initially struggled with emoji display (showing Unicode references) and, even after adapting to text symbols, the game's core "spinning" and winning mechanics appeared glitchy.
10:40 Arithmetic Performance: The model accurately solved a simple word problem involving percentages, demonstrating its capability for basic arithmetic.
11:03 "Steve's PC Repair" Website Test: The model generated a well-designed HTML/CSS/JavaScript website with good aesthetics and interactive elements. Initial issues included a missing main image and an irrelevant secondary image.
13:02 Persona-Driven Website Update: In a compelling follow-up, the model humorously and effectively updated the "Steve's PC Repair" website to reflect "Steve's hatred of Macs." It incorporated aggressive red styling, explicit anti-Mac messaging throughout the site (banners, service descriptions, testimonials, and footer), showcasing its ability to adopt a specific, detailed persona.
16:32 "Johnny's Bitcoin Duplicator" Website Test: The model created a convincing, flashy, casino-themed "Bitcoin duplicator" website, complete with speculative success rates and a private key input. It flawlessly followed instructions to remove all disclaimers about the site being a "demonstration," highlighting its precise instruction following.
18:41 3D First-Person Shooter Game Test (Python): An attempt to generate a simple 3D FPS game in Python using OpenGL and Pygame yielded initial errors but eventually produced a basic navigable 3D environment, though it lacked full game functionality.
20:43 Key Takeaways & Conclusion: The presenter concludes that the Quen 3 Coder 30B A3B Instruct excels in web UI/UX design and text generation with specific personas, but shows less consistent performance in Python game development. Its local run capability and overall impressive web development results make it a highly promising and accessible model.