0:00 Imagine an AI smart enough to operate a robotic arm. It can literally lift a heavy car. It writes complex desktop applications entirely in rest. 0:09 Yet it somehow completely fails to build a simple skateboarding game. Yeah, I mean, it really sounds like a bizarre technological joke. It really does. Welcome to the deep dive. Today we're breaking down some recent developer testing notes. We're 0:23 looking at Alibaba's newly launched flagship model. It's officially known as Quen 4 Max. It was just released on November 16, 2026. We have a fascinating journey ahead of us today. We'll explore its staggering 2 .4 trillion structural 0:37 parameters, we'll decode its wildly inconsistent benchmark testing scores, and we'll watch it code 3D games from one prompt. Well, let's start by looking right under the hood. To understand why it struggles with a simple skateboard, we 0:51 first have to examine its underlying architecture. 0:54 Quen4Max features 2 .4 trillion total parameters, but it only uses 71 billion active parameters per token. Right, and that distinction is incredibly important to understand. It means the model is highly resource efficient. It doesn't use all 1:07 its brainpower at once. It also features a 1 million token context window. It processes text, images, and full video files. It's like having millions of books in a massive library, but you only pull the exact five you need. You pull them 1:23 to answer a very specific question. That happens through its specialized mixture of experts architecture, routing tasks to specialized mini -brains to save computing power. This specific design leads to incredibly polarized benchmark results. Let's 1:37 unpack what those benchmark victories actually mean. On the Deepest WE version 1 .1 benchmark, it scored 75 .7. That is the highest score among all its peers. Exactly. On the agent's last exam test, it scored 46 .7. that easily beats the 1:53 Opus 5 .5 score of 43 .4. And on Bio Mystery Bench, it shows a leading score of 85 .0. But then we see the massive performance losses. Terminal Bench 4 .0 is where things get really rough. 2:07 It scores 48 .5, losing badly to Kimi K 3 .1. 2:12 Kimi scored a massive 67 .2 on that test. And on Program Bench, it only scored a 31 .5. Opus 5 .5 basically doubled that with a 65 .0. These benchmarks test highly rigid programmatic reasoning environments. We also have to discuss its overall 2:30 token efficiency. It achieves better reasoning results while using less power. On Terminal Bench 4 .0, it used about 154 ,000 tokens. The older 3 .8 Max model used nearly 290 ,000. That is a massive leap in operational token efficiency. 2:47 It also dominated the 10 times BPB performance metric. BPB stands for bits per byte in data processing. The code category dropped down to an impressive 1 .67. If it's dominating code efficiency so thoroughly, why is it tanking so 3:04 hard on program bench? Well, theoretical efficiency doesn't always translate perfectly to reality. 3:10 It handles flexible logic beautifully, but breaks down under strict programmatic rules. So it's highly efficient, but fragile outside its comfort zone. Exactly. Benchmarks are just abstract numbers on a computer screen. We really need to see a 3:24 real -world application of this. Does that 1 .67 code score actually translate to reality? 3:30 The source author decided to force a very practical test. They made the AI build a functional 3D environment. It had to build a physics -based 3D skateboarding game. The prompt for this test was incredibly strict. It demanded a fully playable 3:44 3D game from scratch. It had to be built in one single HTML file. It strictly required the 3JS visual rendering library. Absolutely no external game assets were allowed either. It required procedural geometry and materials for the graphics. 4:00 It mapped out explicit player keyboard controls. 4:03 It needed jumping, grinding, and basic directional steering mechanics. I still wrestle with prompt drift myself. Keeping an AI model logically on track is extremely difficult. Yeah, keeping an AI on track is extremely difficult. The prompt 4:19 also demanded complex underlying game physics. 4:22 It needed speed, gravity, jumping, and landing math. Plus, it needed a scoring system for completed tricks. The visual output was absolutely fascinating to examine. It created a colorful, low -poly 3D skate park rendering environment. There were 4:36 geometric ramps, trees, and a city background. 4:39 It successfully generated the required scoring user interface perfectly. It even built a dynamic ollie jump animation entirely from scratch. The rail grind scene is another incredibly nice detail. 4:51 It included visual effects like sparks during rail grinds. But we must emphasize a very crucial caveat here. The screenshots of this game are actually quite beautiful, but they don't prove the game physics actually work. It looks like 5:06 a game, but it might just be a static diorama. 5:09 Right. Visual logic is simply easier for large language models. They mimic visuals better than underlying mathematical physics equations. Are we entering an era where AI builds the facade of software perfectly but leaves the structural 5:22 engineering completely hollow? Yes. Predicting pixels for visual rendering is much easier for an LLM than calculating persistent mathematical physics. Right. It paints a great picture but lacks actual gravity. Physics and continuous momentum were major unresolved question marks. 5:39 The skatepark test left us with some serious doubts. So the source pushed the physics engine even harder. They ran a high -speed car racing test project next. They wanted to test car movement and drifting physics. They wanted to test collision 5:52 physics and track design. The goal was smooth and responsive arcade driving mechanics. The prompt asked for an arcade -style 3D racing game. 6:02 It again required the 3JS visual rendering library. 6:05 It needed a 3D track with turns and straightaways. 6:09 The car controls were explicitly mapped to the keyboard. The space bar was assigned to the handbrake and drifting. The car physics requirements were incredibly dense and complex. It needed speed, acceleration, friction, and braking math logic. 6:23 The racing system also needed a starting countdown timer. It required functional checkpoints placed around the racing track. The visual effects requested were also surprisingly highly detailed. It specifically asked for tire smoke during drift maneuvers. 6:38 The AI actually coded the mathematical logic for tire smoke. It triggered purely off the handbrake drift turning mechanic. Oh, imagine scaling to a billion queries like this. That is an incredible level of nuanced, contextual understanding. It 6:54 linked a mathematical Boolean state to a particle effect. The source shared some important concluding thoughts on this. It's an incredibly strong and cohesive game prototype. It brings the 3D environment and racing mechanics together. This saves a human 7:09 developer hours of tedious manual setup. It drafts the fundamental architecture of the game engine. 7:14 It really does save an enormous amount of time. 7:16 Given how fast it generated these prototypes, does this reduce the value of a human developer's raw coding skill? Not necessarily. While it builds a stunning initial framework, human review and deep debugging remain completely essential for 7:30 actual playability. Yeah, it builds the rough draft, but humans must polish it. Exactly. And building these 3D worlds requires massive raw computational power. Processing a million tokens for a single game is heavy, which brings us to a very practical financial question. What is 7:47 the actual financial catch for you today? Well, we need to break down the two pricing models. 7:52 There's a standard pay as you go API access billing option. It costs $1 .50 per million input tokens. 7:59 Output tokens cost $4 .50 each. Then there's the flat monthly user subscription option that runs about $43 per month for access. This provides a fixed monthly usage allowance for developers. 8:12 Let's look closely at the real -world operational usage data. The entire game testing process barely scratched the subscription limit. It only burned through about 13 % total allowance. That includes building both the skate park and racing games. 8:26 But there's a hidden financial danger mentioned here. Failed model runs still consume your monthly usage allowance. Sometimes the model stalls out during complex reasoning phases. Right, it gets caught in a recursive logical verification cycle. 8:40 It'll just stop after nine minutes, completely dead. But you still pay for that wasted computation time. The system counts those processing attempts against your quota. The final verdict on pricing is fairly straightforward, though. The API is 8:53 for flexible, occasional software development tasks. The subscription is for heavy, predictable code generation experiments. It offers much more predictable costs for heavy developer usage. 9:05 With those failed nine -minute reasoning loops eating up tokens, aren't developers basically paying for the AI's hallucinations? Essentially, yes. Server compute time costs money regardless of whether the model outputs brilliant code or 9:17 just spins its wheels. Basically, you're footing the bill for the machine's daydreams. That's one way to look at it, absolutely. Let's summarize the core overarching takeaway here today. We reviewed the developers' comprehensive notes 9:29 on Quen4Max. Quen4Max is a powerhouse for token -efficient reasoning. It generates impressive 3D game concepts entirely from scratch. But we must remember it is not a magic wand. Building a good -looking user interface is easy for it. 9:43 Making the underlying physics actually work requires a human touch. It struggles with persistent mathematical states and continuous physics. Yeah, you need someone in the loop to review code. They have to test mechanics and fix lingering operational 9:57 bugs. It's highly capable but requires deep human refinement. It is merely a highly capable model for 3D prototypes. Thank you for joining this deep dive with us today. Before you go, I want you to consider something. Models can instantly 10:14 generate amazing visual layouts and basic logic. 10:17 They build any game concept you can possibly imagine. Will the future of gaming value perfect mathematical physics less? Will it value raw, unrestrained human imagination significantly more? If mechanical execution is completely automated 10:34 by future language models, maybe the only thing that truly matters is the prompt. That is a wild and deeply fascinating thought to leave on. Keep learning and stay incredibly curious out there, everyone.
Community discussion