DeepSeek FLASH Destroys Gemini FLASH

Added: 2026-05-10

[00:00:00]Hello, community. So glad that you're back. Yes, today we look at DeepSeek version 4, the flash version, versus the Gemini flash. So, we go here, Gemini 3.1 flash light preview. At the end of the video, I'll show you the non-preview already. I have my standard test, you know it, and here we go. And we compare to DeepSeek version 4 flash thinking.

[00:00:21]So, what will the flash versions do? Are they any better than the other versions?

[00:00:26]And what happens if you go flash, and then you go high? So, let's have a On the left side, you have our old Gemini friend. On the right side, you have the DeepSeek. So, let's see what's happening. As you can see here, in the Gemini version, you do not see a reasoning trace by Gemini. This is a proprietary model, but DeepSeek has an open model here. You see here really the reasoning trace, which is beautiful. And therefore, I do prefer DeepSeek version 4 flash thinking, because you have an idea what the LLM is doing. Now, we have now very fast the first result here, Gemini 3.1.

[00:01:02]And it is coming up, and it tells me, my goodness, this is a bad result. Look at this, how many steps there are. You know, eight is excellent, 10 is okay, but we have here how many steps do we have? We have 20 steps. Are you joking?

[00:01:18]No, we have 14 steps, okay. A less than 20. Okay. But 14 steps is is not really good. No, 3 minutes 30 seconds later, I know you're not really interested in the reasoning trace. DeepSeek also comes to a first result. Now, let's have a result. Hey, look at this, 10.

[00:01:37]You see?

[00:01:38]So, DeepSeek gets a better, shorter reasoning trace, a higher performance, both invoked emergency exit, beautiful.

[00:01:48]So, if you compare the Gemini 14 to the DeepSeek version 4, 10, DeepSeek wins.

[00:01:55]The first win for DeepSeek version 4, flash thinking. Great.

[00:02:00]Now, you know what we have to do. We have to do a verification run. So, here we go. I say verify if your solution is a valid solution that respects all the constraint given here in the instruction of this test. And we will stay live because Gemini 3.1 flash light, it is a flash light and its preview is rather fast. So, let's have a look. You see here on your right-hand side, DeepSeek version 4 real nice is doing here the solution. Yes, the press 10.

[00:02:29]Okay, now verify the final solution. You see everything is clear. You can have a look inside. But look what happens with Gemini 3.1. It tells me my previous solution was invalid. And now I immediately found a corrected solution.

[00:02:43]So, remember we had 14 steps here from Gemini 3.1 before and now we have 15. No, another correction. We are live correcting another time. Now, we have 18.

[00:02:55]My goodness.

[00:02:57]So, we're same time we have DeepSeek finished with 10 button presses confirmed.

[00:03:03]Okay, so I think this is clear. 10 with DeepSeek and now 18 with Gemini 3.1 flash.

[00:03:09]I think the winner is clear. Now, look at this.

[00:03:12]Okay, yeah, floor 29, the emergency exit. This looks a really nice valid solution. DeepSeek second win.

[00:03:21]Okay.

[00:03:22]Now, you know what we're going to say, now.

[00:03:25]If this is here the 10 step, let's have an optimization run.

[00:03:29]And I say try to find a shorter sequence, now.

[00:03:33]Try here to optimize result, maybe try different strategies, search extended, alternate your decision patterns, just find a better solution.

[00:03:42]So, you see, not at all a professional prompt. I just want to see if a normal person handles here our AI system.

[00:03:50]DeepSeek goes off and is trying here.

[00:03:52]You really see that that is trying different patterns, no? It started to say, "Let's try this or let's try this."

[00:03:58]So, it is extending its search space before it is focusing in in on the exploitation of a found solution. Hey, look, they have now nine presses in DeepSeek. With Gemini, we have no idea what is happening, unfortunately. Okay, this will change in the second half of the video, but for the moment we have an optimized result because Gemini finished.

[00:04:18]Gemini finished and Gemini has now a result of What? Have I seen 10 button presses?

[00:04:25]Have I seen 10 button presses with Gemini 3.1 flashlight preview? Yes.

[00:04:30]10 plus the emergency exit. So, this is a good result. This is an average good result. Here, 29 to 50, yes, beautiful.

[00:04:40]Okay, so we'd say, "Yeah, this is a result that we can accept." This is a flashlight.

[00:04:48]Okay, so you might say, "Okay, I know exactly what I get if I use this model here." But now DeepSeek is here also finishing here its optimization and look at this, we have six seven eight Eight. Wow, this is Wow, this is really much better. This is excellent. 10 with Gemini, but eight a shorter sequence.

[00:05:11]DeepSeek has third win. Absolutely, DeepSeek version four highly recommended. I will use this for my highly complicated scientific task if I want to have this. But you know, out of preview May 8 tell us here now available for enterprise. So, here we go, we have here the old uh the new flashlight and we go here with this light. Thinking level is now high. I want to see, if I use this and I put the thinking on high, is a flash as good as something else?

[00:05:44]So, we We now a benchmark by DeepSeek version 4 flash of eight.

[00:05:49]Okay, it took some minutes. Okay, the time you have to give it time. But what happens if I set the thinking level on high and I go with flash light?

[00:05:58]Gemini 3.1 You see, this is not a real thinking trace. This is here a summarized artificial synthetic reasoning trace to, yeah, calm me down, but I don't like this. This is not really this is here not the pure thing. This is a synthetic optimization. So, okay, at least we get something. If you work directly on the sandbox, on the playground here, on Gemini, beautiful. So, finalizing the strategy, beautiful, confirming here energy balance, validating here the code sequences.

[00:06:31]There are multiple interlinked optimization to be done.

[00:06:34]The first result now with flash light high out of preview is, my goodness, what is this? This is This is bad. This is real bad. 20 presses. This is the maximum of presses allowed.

[00:06:48]No, not acceptable. So, optimization run.

[00:06:52]Flash light high That's a real bad result. Try to improve your result, find a shorter sequence of button presses, go.

[00:06:58]My goodness.

[00:07:00]Mapping the early floors, [clears throat] evaluating alternative path jump potential, revising the jump logic.

[00:07:08]You see, this is just marketing slang.

[00:07:10]This is not really that we have a look into what is happening in the reasoning traces. So, therefore, I like DeepSeek and I like open system that I really see something that sounds like it could be the real reasoning trace.

[00:07:23]Okay, Gemini 3.1 is now here flash and light maybe not the best model to apply here to a scientific target.

[00:07:34]Even if you go with a thinking level high, this was really quite an eye-opener for me. So, but let's let's do it again. Let's have a look. Can it find a better solution? Now we tell it, "Hey, optimize yourself."

[00:07:49]Clarify here the end [clears throat] game strategy. This is a strange wording.

[00:07:54]Refining the target floor, the teleport path, beautiful.

[00:07:57]Finalizing. Oh, here we have it now. 12.

[00:08:00]Now we have something. Optimization result is now 12 button presses.

[00:08:05]This is less than average, but okay, at least it is not 20 now. So, no, thinking level on max is not a solution if you have a flash and then even a light mile here, even Gemini is not able to perform a task here. No way, this is incorrect.

[00:08:23]I'm so sorry. This is not a correct solution.