This video provides a first-look, hands-on test of the newly released Quen 3 coder AI model, described as a "strong coding agent capable of long horizon tasks." While official technical details like a model card are not yet available, the presenter leverages community information indicating a potential context length exceeding one million tokens and a Mixture of Experts architecture with 480 billion total parameters (35 billion active). The evaluation focuses on the model's code generation capabilities, particularly for web-based applications using HTML, JavaScript, and CSS.
The tests reveal exceptionally fast token generation and a significant improvement in UI/frontend design compared to previous models. Quen 3 coder successfully generates a highly interactive web-based operating system, aesthetically pleasing business websites, and a visually impressive startup landing page incorporating psychological marketing elements. While some complex outputs like a racing game and an FPS exhibit functional limitations (e.g., calculator arithmetic errors, lack of mouse aim in FPS), the model demonstrates strong creativity in adapting content from "The Great Gatsby" into a platformer game. The overall impression is one of marked advancement in UI generation and the ability to produce functional, complex web applications from varied prompts, despite some minor bugs.
Testing the Quen 3 Coder: A Deep Dive into its Web Development Prowess
0:00 Introduction to Quen 3 Coder: The video introduces the newly released Quen 3 coder, an AI model for code generation, eagerly anticipated by the community.
0:48 Initial Model Information: Based on unofficial sources, Quen 3 coder is a Mixture of Experts (MoE) model with 480 billion parameters (35 billion active) and a reported context length exceeding one million tokens.
1:45 Fast Token Generation: A key initial observation is the model's significantly faster token generation speed compared to previous Quen models.
1:36 Web-Based Operating System (First Test): The model generates a web-based OS using HTML, JavaScript, and CSS.
2:06 Single-File Output: The entire code is generated in a single HTML script, which is less ideal for production but common for AI-generated code.
2:22 Impressive UI/UX: The OS features resizable, translucent windows, a functional terminal (with commands like help, whoami, about, date, clear), a text editor, and a file explorer, indicating strong UI/frontend capabilities.
4:34 Calculator Bug: A minor functional error is noted in the calculator, where arithmetic operations reset the display without calculating.
6:35 Web-Based 3D Racing Game: A more challenging prompt for a low-poly, first-person racing game with an opponent and basic controls.
7:30 Aesthetic and Controls: The game has a retro aesthetic and a thorough controls menu.
7:53 Functional Limitations: The game is slow, and the gameplay logic is not fully robust, though the pause/restart functionality works well.
9:09 Iterative Improvement: The model shows slight improvement in a second attempt after receiving critical feedback, indicating an ability to refine its outputs.
10:17 Platformer Game from "The Great Gatsby": A test of context length and creativity, asking the model to generate a platformer game based on the entire text of "The Great Gatsby."
11:01 Creative Adaptation: The game effectively incorporates themes from the book (e.g., collecting "golden opportunities," reaching the "green light").
11:31 Jump Logic Issues: The game is functional but has issues with jump physics.
11:52 Simple Website (Steve's PC Repair): The model generates a website for a fictional PC repair business.
12:00 Strong Business Aesthetics: Features include strategically placed yellow buttons for calls to action, an appropriate background image, hover effects on service cards, relevant emojis, and well-structured "About Us" and "Testimonials" sections.
13:49 Properly Done Footer: The website includes a well-designed footer with quick links, correct social media icons, and contact information.
14:37 Startup Page (GPU for Rent SaaS): A prompt for a startup landing page, requiring functional and aesthetic design with psychological and subliminal marketing elements.
15:06 Advanced UI Elements: Incorporates color psychology, floating elements, subtle animations, gradient text, rounded corners, and layered backgrounds.
15:31 Impressive Hover Effects: Notably sophisticated hover effects on navigation links are observed.
15:56 Detailed Service Tiers: Offers clear pricing tiers for GPU rental, including specific GPU models (e.g., Nvidia V100, A100) and system details.
17:21 Contextually Appropriate Social Icons: Social media icons (Discord, GitHub) are tailored to the tech startup context, showing contextual awareness.
17:51 Web-Based Music Sequencer: A simple prompt to generate a music sequencer.
18:11 Excellent UI Design: The model produces a visually appealing and intuitive user interface for the sequencer, with appropriate instrument emojis.
18:25 Significant UI Improvement: The presenter highlights this model's substantial leap in UI generation capabilities.
18:58 Web-Based FPS (Neon Strike): A vague prompt for a first-person shooter game.
19:36 Functional Game Elements: Generates a game with a health bar, working ammunition counter, and a functional shooting mechanism.
20:21 Mouse Aim Issue: The primary functional flaw is the lack of mouse-controlled aim.
21:03 Impressive Output for Vague Prompt: Despite the aim issue, the complexity and functionality generated from such a vague prompt are deemed impressive for an open-source model.
21:38 Benchmarking Expectations: The presenter anticipates Quen 3 coder will benchmark very highly against other open-source coding models.
22:04 Local Run Requirements: Running the model locally will require a "beefy system," potentially a DDR5 heavy RAM build, due to its large parameter count.
22:38 Overall Impression: The model is "very impressive," particularly in its ability to generate excellent user interfaces and produce creative, functional code, serving as a strong "taste test" of its capabilities.
Error: value error Invalid operation: The response.text quick accessor requires the response to contain a valid Part, but none were returned. The candidate's finish_reason is 1.