Step 5 Preview: IA barata e SURPREENDENTE programando

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Step 5 Preview: IA barata e SURPREENDENTE programando · Channel: AI ProgBr · 22:42 · Watch on YouTube

Topics: AI evaluation and benchmarks, Software engineering with AI, AI content generation, AI models, Local and open-source models

Summary

The presenter introduces Step 5 Preview, describing its size, context window, vision capabilities, pricing, and benchmark results. He then tests the model by generating and playing Doom, Mario Kart, Resident Evil, and Crash Bandicoot games, noting both successful features and bugs. Resident Evil is praised as the best version tested so far, while Doom becomes playable after a fix and the other games show mixed results. The video closes with comments on the model’s planned open-source release and its potential use for game creation.

Key points

Step 5 Preview is described as a 600-billion-parameter model with 27 billion active parameters, a 1-million-token context window, and vision capabilities.
The presenter reports that the model scores 44 on Artificial Analysis’s intelligence index and is competitively priced compared with several other models.
The video compares Step 5 Preview’s benchmark performance with GLM 5.3 and Kimi models, noting that it remains competitive.
The presenter uses a generic command and prompt to generate a Mario Kart-style game with Three.js.
The Mario Kart game is considered a good result for a basic prompt, though its controls are inverted and the character selection is unavailable.
The Resident Evil game includes functioning movement, combat, a typewriter save feature, keys, and collisions, but has visibility and aiming issues.
The Crash Bandicoot game has jumping, spinning, collectibles, and multiple levels, though later levels look much the same.
Doom initially has a door issue, which the model eventually fixes; the game includes weapons, enemies, and multiple levels.
The presenter says Step 5 is planned to go open source on October 15 and considers whether a quantized version could run locally.

Chapters

Start	Chapter	Summary
00:01	Introducing Step 5 Preview	The presenter introduces Step 5 Preview and discusses its parameters, context window, vision capabilities, intelligence score, pricing, and benchmark comparisons.
03:47	Testing Doom and Mario Kart	The presenter begins testing generated games, finding a door issue in Doom and then trying a basic Mario Kart-style game with inverted controls.
09:58	Resident Evil gameplay	The presenter explores the generated Resident Evil game, testing its save feature, combat, keys, and collisions while encountering aiming and visibility problems.
14:04	Crash Bandicoot and game progression	The Crash game offers jumping, spinning, collectibles, and several levels, although the later levels appear visually unchanged.
17:41	Returning to Doom	After a fix, Doom becomes playable, with the presenter testing its doors, enemies, weapons, map, and arena levels.
21:33	Final impressions and open-source plans	The presenter sums up Step 5 Preview as useful for making games, mentions its planned open-source release, and invites viewers to comment.

Quotes

a 600-billion parameter model, which has 27 billion active parameters. — 00:32

It turned out pretty basic, but that’s exactly what I asked for. — 08:20

Best Resident Evil we’ve seen so far — 13:30

another very interesting model that you can actually use, at least for making games. — 21:33

Concepts and entities

Step 5 Preview, Step Fun, Artificial Analysis, Qwen3, GLM-5.3, GPT-6 Astro, Claude Opus 5, Three.js, DOOM, Resident Evil, Mario Kart, Crash Bandicoot, APRA Devis

Full transcript (original language)

00:01 Sensational. The best Resident Evil we’ve seen so far. The channel now has hardware, image, and video content, and today I’m bringing back our games, because I know some of you have been missing them. And to help me with this task, I’m bringing in Step 5, which is in preview version, and is the new model from Step Fun. Actually, I really like Step Fun’s models. I had Step 3.5 Flash running here on my Mac, it was very good. Then came 3.7, now we’re on version 5, and this model is a bit larger than the Flash one. Let’s head

00:32 to the screen so we can learn a little more. Step 5 Preview, a 600-billion parameter model, which has 27 billion active parameters. It’s a model with a 1-million token context window and vision capabilities as well. The first thing we want to look at here is this model’s intelligence. So, if we go here to Artificial Analysis, we’ll be able to check some of that out. And intelligence in general, right? On the Artificial Analysis intelligence index, it sits here with 44 points. This places it between GLM 5.3 and Qwen 3.

01:02 To give you an idea, Qwen 3 has 2.8 trillion parameters, okay? Step 5 Preview has 600. This is what’s been happening in the industry. We have smaller models that are getting better and better. Today, Qwen 3 827B is considered to be on the same level as Opus 4. Look how crazy that is. A model that I can run in FP8 here on my machine. By the way, I built a new machine with two ARC B770s. It had a few issues there. If you want to know a little more, I’ll leave the video here in the card. But we can run a Qwen 3

01:33 827B on it in FP8, with almost no loss at all. So, models have been getting better and smaller. And that’s the case with the Step 5 Preview as well. Now, another interesting thing about this model is the price. We’re talking here about 26 cents per 1 million input tokens and 2.70 per 1 million output tokens. Cached reading comes out to a few cents, and again, this model has vision capacity for images and video. If we take the models that are close to it, GLM 5.3 costs 1.40 per 1 million input tokens and 4.40 per 1 million

02:08 output tokens. In this case, it costs 26 cents. As for the Kim K3, which actually sits behind Step, it will cost 3 per million input tokens and 15 per million output tokens. I’m hopeful that the next Moonshot model will be a bit cheaper. That’s already the case for Kimica 2.7, which for coding tasks ends up being comparable to Kimica 3, but at a much lower price. That would be 95 cents per million input tokens and 4 per million output tokens. Even so,

02:39 it’s more expensive than Step 5. When we look at the benchmarks compared to other models, we see that it is a truly competitive model. So in Deep SWE version 1.1, which is one of the important benchmarks today, it scores 67.7, above Kimica 3 and above GLM 5.3. It falls behind, quite far behind even, GPT-6 Astro and Claude Opus 5, but that is expected. In the Step Code BCH, the gap is even larger in relation to Kim K3 and GLM 5.3. In the terminal BCH, it appears here between GLM 5.3, which was

03:14 well ahead of it, and Kimica 3. But that’s it, it’s always in the game against these two models. Remember, GLM 5.3 is a model with over 700 billion parameters. It’s not that big of a gap. Step 5 has 600 billion, but Step 5 is smaller and at least according to the benchmarks, it performs better in most cases. But let’s get to what matters. We are here to see games. So let’s see games. I’ve prepared four little games here today for us. The first is a classic Doom that we’ve also done on other models. So you can get an idea here. I also brought back our Resident Evil. I think in one of

03:47 the last videos I forgot to tell it to use Three.js. And then we got an interesting version, even top-down, but a little different from what I expected . This time I didn’t make that mistake and brought other games. I brought back an old-school Mario Kart, it’s been a long time since we’ve done that, and I brought Crash. Let’s start with the classic Doom and we’ll go from there. Let’s go. We have our Doom here. The first test is to see if it runs or not, right? I’m seeing some things here in the back, which is a good sign. So, I think it will work. Let’s go. Since this Doom prompt is well-defined, the

04:17 expectation is that it works like the other Dooms we’ve seen here in recent videos. Woah! Man, look at the movement of the weapon when I turn. That’s pretty interesting, right? Let me grab this here. Ammo. Look at this. We have a door, look. I can already see he made a mistake I’ve been seeing, which is this door facing the wrong way, right? The door should be closing here, and it’s working. It’s closed here, but it’s like this, in the other orientation. Let’s go. Oops, what

04:49 happened here? This here is the chainsaw. Man, but the chainsaw shoots. Strange. And it looks like I have ammo for it there, see? Do you see it down there? 103. Well, let’s see here first. Let me see if sprinting is working. What happened? Uh oh, something got stuck here, look. Sprinting is working. Okay. How do I open the door? It’s not the Is that one locked? No, it shouldn’t be. And wait a second. Is it me who’s confused? It is me. Ah, what a

05:26 shame. It was so cool. Man, the door isn’t opening. I’ll ask him to adjust it. While he adjusts it, let’s take a peek at our Mario Kart. Man, it looks cool back here, huh? So, let’s see here . It’s the It uses the item. Okay. Space, drift. Man, space to use the item would be easier, right? Having a joystick on the screen. Oh, by the way, will our joystick here work, look? I don’t think so, huh? Oh, the joystick is working. But the accelerate Ah, okay

06:04 . It’s because I wasn’t on the track and it’s reversed. I turn left, it goes to the right. And why is the turbo there sideways? But man, it’s been a long time since I’ve done Mario Kart here, okay? So, considering that, I don’t know if it’s a Gosh, I can’t play and talk. Ah! Ah! Learned here how to

06:38 use the I was trying to figure out which button used the item. There, now we’re talking. Now, now, now I’m going to win. If Man, but the turtle is slower than Makes sense, right? The turtle is slower than the kart, man. It’s not going to hit anyone here. I have to shoot from very close or just leave it like this, like a trap there.

07:11 Man, playing with reversed controls is really hard. Or am I just making excuses, right? That could be it too. T. Look, I’m in first place, huh? That’s it, a bit weak there, but that’s okay. Look, now I’m going to grab the first one here. Then I’ll use mine. T t t. I’m dropping the banana now, man.

07:49 It’s much better than the other Mario Karts we made here, tad.eesxe browser.js, but that could just be a general evolution of the models, right? It’s been a long time since I made a Mario Kart. I’m going to start making them again, to help us understand how everything is evolving, because, overall, I thought this result here was very good, considering it still makes the mistake of inverting controls, but overall, I thought it was a very interesting result, man. I can’t choose here. Can I? Where is it? Let me go to

08:20 the menu. I can’t choose my character. It turned out pretty basic, but that’s exactly what I asked for. I didn’t give many explanations, I didn’t ask for a Mario Kart clone, and this is what it delivered to us. By the way, to create this Mario Kart clone, I have a sort of generic prompt, okay? That I put inside a command. So I run my “create game” command here, and then I pass which game I want to create and which technology I want to build it with. In this case, it would be something like Mario Kart and Three.js. Done, that was basically what I did, and the result we saw was based on that. But when we

08:54 press enter here—I’ll press enter and then cancel. Just so we can see here, look. What we do is generate this entire prompt, replacing certain keywords with Mario Kart and Three.js. But our base prompt is ready. It was placed inside a command that receives arguments and adjusts the prompt according to those arguments. This is actually something I teach in my APRA Devis course. If you are a programmer who is dependent on the biggest models in the world and consequently the most expensive ones, and it’s getting too expensive for you to develop with AI,

09:25 this course is for you. If you are a programmer who hasn’t dived headfirst into this new way of creating software with AI yet, this course is for you. If you come from any other field, but you want to learn how to use agents better or you want to create your own applications, even if they aren’t commercial things that you’re going to launch, sell, or do professionally, this course is for you too, okay? We made a course focused more on devs, but even if you’re a vibe coder or not a dev at all, you can reach a new level too if you learn everything we teach there. So, I’ll leave the link down

09:58 below, take a look, sign up, and I’ll see you inside. Let’s see if it finished our Doom here yet. It hasn’t finished yet, it’s still adjusting our Doom over there. I thought this adjustment would be simple, but let’s let it keep running. In the meantime, we’ll check out our Resident Evil. Whoa , Resident Evil. Look, our joystick works here too. Okay, let’s go. Let’s start the game here. I don’t even know what to expect from this, okay? But let’s go. Whoa! What is that? Where is it? How do I How do I do anything here? Ah, take that. Wait, which way am I facing? Oh no. How do you run? Ah, he’s

10:38 he’s only shooting upwards, man. He’s not shooting where I’m facing. See? He’s shooting upwards. Ah, then take that. Okay, it’s got a small bug, but okay. Man, this is really cool. What is this thing here? It’s a typewriter. Save. How do I do this here? I’m typing . Whoa, I saved the progress, huh.

11:13 Where is it? I don’t know which button does the thing here. Why am I going this way? Locked door. But I just passed through the locked door, right? What happened? Damn, man, that’s crazy. Oh, I’m stuck here on the box. Man, the collision is good. Let’s head over here . Ah, take that, take that. Hey, it didn’t die, it didn’t die, it’s not dying. Ah, it died. Now I’ve got the hang of it, see? I have to position myself so he stays in front of me. If

11:47 he corners me here, I’m screwed, right? If he stays here like this, I’m going to die. There’s no way around it. Let’s go. Oh, I used one there. It sounds like he farted, doesn’t it? Ah, but it’s using. Oh, now take that, take that, take that, take that. All right. Here, you have to do this. You have to come this way. And take that, take that , take that, take that. Nice. Now I’m I’m I’m getting the hang of it. I’m getting it. Grab that thing over there. What is that thing over there? It’s a key. Emblem. Got it. I got the emblem.

12:20 How do I open the door here, look? Opened it. I’m screwed. I’m screwed. I’m screwed. Take that, take that, take that, take that, take that. Where’s my ammo? Where’s my ammo? Where is it? Is it here? No, it’s here. I don’t know. I don’t know how many bullets I have. I have 14 bullets, but I’m using the shotgun. I think I’m using the shotgun. Man, I picked up the ammo there, but nothing changed. I don’t know how to change it. Was there anything else there? I don’t see it. Let me get out

12:56 of here. I’ll stay up here more. Let me see if I can look through the corner. Man, I left that screen and came back, and the zombie is still here, man. Interesting. Look, I came over here. But there’s nothing here, right? Okay. Now we’ll be able to open that door over there, see? You see? The zombie is still here, man. How crazy. Come on, let’s go in there now that we have the

13:30 key. Whoa, I can’t see anything, you know? He cornered me there, I’m screwed . Game over. Man, the only mistake, the only, the only flaw was that. I couldn’t see anything there, man. Sensational. Best Resident Evil we’ve seen so far, okay? I’ll bring more of these prompts, but so far this is the best one. Too awesome. And here we have

14:04 him, Crash Bandicoot, Lost Island. Run through the jungle, collect the Wumpas and spin into the boxes. Awesome. Let’s go. Whoa. I got hit there, man. He does a double jump. I’ll be grabbing the stuff there. There’s the spin. If I spin here, it works. Ah, I’ll try to spin into that little creature there.

14:35 But the game is playing by itself here for me. It wasn’t supposed to be like this. Was it? It wasn’t. It wasn’t. Look here. Whoa. What happened? Game over. Why? Let’s go again. Let’s go again. Now I’ll collect as much as I can. I’ll be breaking the boxes here. I’ll be jumping here, look. Come on, come on, come on. Whoa! Wait, where are the enemies now? Gee, that looks bad,

15:12 huh? It’s bad. It’s bad. It’s bad. I think I have to double jump there. But visually it doesn’t look, it doesn’t look like I need to, right? So I’ll I’ll make sure, but I didn’t think I needed to there. But let’s see. Here. I’ll skip that too. There’s something on the floor I don’t know what it is. Oh, now it’s working. There we go. Let’s see if that’s really it again.

15:44 Take that double jump. Now it worked. Now I went far. Take that. Now I’ll always go with double jumps. Whoa, that one was dangerous. I thought I was going to mess up. Where’s level one? Oh , it’s it’s right there. We’re getting there. We’re reaching the end. Oh, man. And I can jump and spin at the same time. And I reached the end of level one, I think. I beat it. I beat it.

16:18 Next level. But what changed here? Nothing changed, did it? Same thing. Okay. Visually there’s no difference at all. Except I’m on level two now, supposedly. I wonder if there will be new enemies? Here is oh, he jumps and spins. Oh, wait, I went right over him. That’s not right. I have to make sure I hit him then. Whoa. It’s getting faster , huh? Where is it? And look, level two

17:08 is gonna end there and then we’ll see if there’s anything else. Level three. Level three. But it’s all the same. There weren’t any new enemies, nothing different at all. But cool, cool. Interesting result. Let me stop here. Cool interface too, everything is nice. And look at our Doom, it’s running some tests here, it seems. Let me expand this here. We had one failure, right, it couldn’t get through the open doors, but I think it’s almost solving it now. Let’s wait just a little longer. Oh, it passed, it passed now. It passed. I

17:41 think it ends now. I think now it’s going to let us play. Awesome, it finished here. Let’s run it again. Our Doom is also working here with the joystick. You can even play like this, but this joystick I have, that I’m using today, doesn’t have an analog stick, it’ll be harder to play. Let’s play with the mouse and keyboard instead. Let me grab this ammo already. First thing I want to see is if Ah, I could have asked it to fix that door issue there, right? But that’s okay. Let’s try to enter this door here. Oh, the door opened. Awesome. And take that . Take that. Oh, it’s kind of a robot

18:15 this time. Cool. Different. So, look, all the doors here on the south facade, I don’t know if it’s the south facade, they’re oriented sideways, but let’s go . Take that, take that, take that. Where’s the Let me, let me go with the chainsaw. Let me try to kill with the chainsaw directly from afar. Doesn’t work, right? Clearly doesn’t work. You see that our enemies here are these cardboard cutouts, right? But this is expected, okay? It was supposed to be

18:48 like this. Let’s go. Whoa, wait, wait, wait. Let me. Where’s my mouse? I’m struggling here because I can’t, I can’t see the map. So here we have the map. Down here we have the weapons we have. We still don’t have a shotgun or rocket launcher, but we’ll solve that right now. Here, look. Here we already solved it with the shotgun. We already have the shotgun in hand, okay? And we’ll be able to come back here. I think the level is a bit too dark for

19:23 my taste, but that’s okay. Do I already have the key or not? Ah, I already have the key. Whoa, look. Man, this weapon is tough here, isn’t it? Uh oh, there’s another one there. We’re screwed. Let me try with the pistol. Man, but the pistol ran out of ammo. Oh, no, I was using the shotgun. Pistol is here. Now the shotgun ran out of ammo. Pistol is faster, right? But the shot is also weaker. But that’s okay, it worked. Nice. Let me get more health there. Ah,

19:57 I’m already at full health. That little shield that came, it came in handy, look. Grab another shield here. Awesome . Now I’m going to level two. Top. Level two. Final arena. Let’s go. The arena is different from the other one. It’s a little different, right? I think I’ll come here first. Whoa, let me switch to the shotgun. No, because I’m out of Ah, this one shoots now. Man, I

20:28 think he’s too far and the bullet isn’t hitting him right, man. I grabbed shotgun ammo, right? I can go with the shotgun now. Let’s go here. Here we’ll just grab more ammo. Got it. Good. And here I think I’ll enter here. I’ll try to get the rocket launcher right away, it’s right there, look. There’s the rocket launcher. I’m going for him. Done. Take that. There. This one took just one. Boom. Now we’re talking. The rocket launcher will make my life

21:02 easier here. On this big one. He needs two shots. I already hit this one. Take that. Man, I’m getting good at this. Or maybe this version is easy, I don’t know. Ah, there are two more here. I hadn’t seen them. That messed me up. Has to be the rocket launcher. As soon as I celebrated. I almost got screwed, man. Grab everything here, look. Alright. Nothing left to grab. Awesome.

21:33 I’m heading out. Victory, cleared. Top notch. Turned out good, man. Turned out good. Well, my friends, another very interesting model that you can actually use, at least for making games. It performed well. Go. Also, on October 15th, this model will go open source. 600 million parameters is a bit hard to run locally, but it’s always good to see these models go open source, especially because, depending on the license, other providers can host it. At even cheaper prices. I’ll see if we can run a Q1, Q2 version, maybe locally

22:09 here. But even Q1, Q2, since it has 27 billion active parameters, I think it will be a bit slow. Anyway, we’ll see. And I want to know what you thought of this model. Leave it down below in the comments. If you want to see a video where I create a virtual influencer, take a look at this video here. Now, if you want to see a video about the new machine I’m building with two RTX 4070s , take a look at this video here. If you want to learn more about developing with AI, look, it’s right here, this link has all the information you need to make a decision and join our course. That’s it, I’m signing off, until the

22:42 next game. We’re in this together.