DeepSeek V4 Lite LEAKED + GPT 5.3 (Garlic) This Week?

Two leaks have just dropped and both of them are worth your attention. First, what appears to be Deepseek version 4 light and the outputs are turning heads for what reasons most people are getting wrong. Then GPT 5. 3 code name garlic with the release date and a benchmark number that if accurate suggests Open AI is about to have a very good week. So, let's get into it.

We might have our first real look at Deepseek version 4 light. A post from Marmaduke091 is making the rounds and they're claiming these are outputs from DeepSeek version 4 Light, a model nobody has officially announced yet, but are expecting. They're being honest about the uncertainty. They say take it with a grain of salt, but they also say they don't know any other model that can do what they're showing. And after looking at it, I think they have a point. The comparison is version 4 light with thinking off versus version 3. 2 with thinking on. two prompts, an Xbox controller, and a pelican on a bicycle. Version 4 Light wins visually and pretty comfortably. But that's not the story. The story is what these outputs are actually are. These aren't AI generated images. Deepseek version 4 Light is apparently writing SVG code, vector markup, and what you're seeing is the render result of that code, and the line counts are sitting right there in the post. 54 lines for the Xbox controller, 42 lines for the Pelican. That's the number you need to hold on to. SVG is unforgiving. It's not a diffusion model blending pixels until something looks approximately right. It's explicit geometric instructions, pads, coordinate, shapes, layers. You describe the object correctly in code or the output falls apart. There's no smoothing, no approximation. And an Xbox controller is not a simple object. Asymmetric body, two joysticks, a D-pad, face buttons, triggers, all of that. Spatially accurate in 54 lines is also amazing. That's a real constraint to work inside of. And version 4 light is doing it with thinking turned off. No chain of thought, no extended reasoning pass, just direct output. Version 3. 2 with thinking enabled produces something visibly messier, which once you think about it makes sense. Reasoning tokens help you catch errors in logic, verify steps in proof, audit a calculation. They don't help you understand where to place a curve or things like that. Either the model has spatial structure baked into his weights or it doesn't and no amount of verbal scaffolding closes that gap if it isn't there. What this suggests about version 4 light if these outputs are genuine is that something meaningful changed in the base model not in the reasoning layer on top in the foundation the efficiency alone points to that 42 lines for a pelican on a bicycle isn't just clean output it's economical output most models asked to generate SVG either produce bloated redundant markup that technically renders or produce something that parses but looks nothing like the prompt getting complex geometry down to the line count while keeping it accurate implies the model has learned something real about how to represent objects, not just how to approximate them. And that has implications well beyond SVG code quality, spatial reasoning, structured output. If version 4 light is doing this on visual task, you have to wonder what it's going to do on everything else. We don't have official confirmation. We don't have model specs, training details, or a release date at this point, but the claim is specific and testable. Someone with access could reproduce or debunk this quickly. and Deepseek has made a habit of showing up faster and leaner than anyone expected. Version 4 light might be the next one.

So, what I'm going to show you guys are three examples. One from the Deepseek version 4 that we saw, but the actual code line. So, you can see that it's actually 54, then from Opus 4. 6 and then from Gemini 3. 1. So, what you're looking at right now is the Deepseek one, the Deepseek version for light. And it is actually 52 lines of code as the tweet mentioned. So, he isn't lying about that. This is actually 52 lines and it's pretty I'm like obviously it doesn't look 100% like a Xbox controller but we're going to start appreciating the 52 lines and like what it looks like real soon because I asked Cloud Opus 4. 6 as well. So what you're looking at right now is an Xbox controller that Cloud Opus 4. 6 tried to create and yeah you can tell like this does not look like an Xbox controller and for some reason I don't know why the model decided to add some additional features. So, of course, the lines of code are way more compared to 52. I think it's closer to how much is there? We're at 439 lines of code, but then there's obviously the index file. Then there's a separate different CSS. I think the CSS only has there. But like overall, I think it obviously has more code, but I'm guessing it's because it tried to add these additional features, but obviously the Xbox controller does not look like it. And then I asked Gemini to create me a SVG controller as well. And then I pasted it into an XML viewer and I saw it has 54. So Deepseek beats it by about two lines I would say. But anyways, this controller out of all three of the outputs is definitely better and we kind of know this because the 3. 1 model if we have looked at the generations when it was being leaked as well. We saw that it was better at SVG generation. So we can see that. But if we're getting a Deepseek version for a light and it's able to be as efficient as the newest models from Google from, you know, Opus 4. 6 and produce something that's competitive like we saw the Xbox controller, which looks pretty similar to the one Gemini 3. 1 created, Deepseek version 4 Light might once again shake up the whole market. All right, we have

a release date and some numbers for GPT 5. 3 Garlic. And this post is worth going through carefully because it's the most concrete thing we've seen yet. Dan is claiming Garlic drops Thursday, February 26th, or also known as the 5. 3 model 4 days from now. And he's not just dropping a date. He's attaching a simple bench leaderboard to back up why this is a big deal. So, let's talk about simple bench for a second because the benchmark you use matters. A lot of AI leaderboards have reliability problem. The models get trained or fine-tuned on data that overlaps with the benchmark. scores inflate and the number stops meaning what it's supposed to mean. Simple bench is specifically designed to resist that. It tests common sense reasoning. The kind of grounded practical understanding of how the world works that's surprisingly hard to fake. It's not trivia or math is whether the model actually grasps how things behave in reality. And here's a leaderboard as it stands right now. The human baseline is 83. 7% and the best model currently on the board is Gemini 3. 1 Pro Preview at 79. 6%. 6% which notably is still below humans. Gemini 3 Pro comes in at 76. 4%. Cloud Opus 4. 6 is at 67. 6% and Gemini 2. 5 Pro and Cloud Opus 4. 5 are both sitting around 62%. GBT 5 Pro is at 61. 6%. Every current Frontier model is below human baseline on this benchmark. That's the context. Dan is saying that GPT 5. 3 will clear that 83. 7% surpassing human baseline on simple bench. And he goes further saying that it blows every previous model out of the water on all non-coding benchmarks. The word he uses is huge leap. His framing is a GPT3 to GPT4 moment, which is not a phrase people throw around lightly. That was a generational jump and not an iteration. His explanation for how OpenAI got here points to two things. First, they had the best RL and post-training pipeline in the industry for a while. That goes back to 01 when they were the first lab to seriously apply reinforcement learning to inference time reasoning. That work has been compounding. And second, he's suggesting that advantage has now extended into pre-training as well, citing comments from Mark Chen on Ashley V's podcast as a signal that something significant changed upstream. Sam Alman's public comments apparently point the same direction. Dan reads them as confirming as major progress and not just incremental progress. And his closing take, this may be deserving of a major version bump, meaning GPT 5. 3 might actually be GPT6 territory in terms of capability, but just shipping under a different name. That's a prediction, not a confirmed fact, but it's coming from someone who seems to have visibility into the pre-release picture. Thursday is 4 days away. If the simple bench number is real and if the non-coding benchmarks claims hold up, this is going to land as a genuine step change and not just a simple model update. So, Garlic February 26 will know

soon enough. But that's it for today's video. Make sure to subscribe to the channel, follow us on Twitter, follow the world of AI, and don't forget to subscribe to our newsletter. We post constantly, and you don't want to miss this. I'll see you guys in the next