**Julian Goldie** (0:00)
DeepSeek V4 Flash just came out of preview. There's a brand new version, and it's a giant leap. I actually tested it out, so you don't need to. And I want to show you what I built, over 50 builds with it. So I'll show you side by side how that compares against DeepSeek V4 Pro, so you can see the huge leap and the huge difference. Now, if we go straight to the builds over here, we can see what we've created and how we're side by side. So the first thing that I would say is, like, out of everything I've tested, DeepSeek V4 Flash 731 is number one, way, way better than the original preview. Number two, it is even better than the Pro version. So the Pro version is on the right-hand side over here, this version right here, and it doesn't really do anything in this particular build. Whereas, for example, we look at DeepSeek 731 and look at this, like, it creates some amazing stuff. And I actually can't wait for the new Pro version to come out, which is coming out soon. That's going to be even better. That could be like Fable 5 level technology, but open source, like, how crazy would that be? So who knows what's going to happen next? We actually did build some stuff that didn't work so well.
But I'll come on to that in a second. This is another game that we built over here, and it's just way easier to control and nicer to control than the DeepSeek V4 Pro version. And it can create, like, some really interesting builds and some 3D games. And, you know, some people say, oh, you can't code with DeepSeek V4 Flash. You absolutely can, as you can see right here. And if you want to check out these builds, you can check it at Goldie Bench. Now, if we have a look, for example, at the DeepSeek V4 Pro version, like, I can't even view what we're creating here, right? Like, it's the camera angle and everything else is pretty bad over here. So most of the builds were like this. They were just way nicer with V4 Flash overall. Now, if we have a look at this one, for example, and this is actually something that surprised me as well, was like, when I was using this, it actually created a lot of stuff that was better than some of the frontier models I created. Now, I'm not saying this is better than frontier models, it isn't. It's not better than Fable 5 or GPT-516, but my point is, it can hold its own. It can actually build useful stuff, it can build interesting stuff. And when I built out and coded out games like this, they were smooth, they were easy to use, and it actually worked, which quite often, these sort of builds, especially if it's complex, if it's 3D, if it's a game like this, usually they don't work so well. So that was interesting as well. I mean, if we compare this side by side, look at the difference in graphics. This feels a lot more like sort of basic in terms of the 3D effects and everything, whereas DeepSeek V4 Flash definitely did a better job overall as well.
Now, if we have a look here, we've got some other games that we created. It was just create some nice things. Now, it's not always like this. So if we have a look at this one, for example, it's kind of buggy, like you can see through the floor, which doesn't make any sense. It looks like it's floating, which it shouldn't be. And also the graphics here are pretty basic, right? It's not that interesting, along with the controls. So it does fail sometimes. Now, what DeepSeek are actually trying to do with this model is they are trying to build something that's designed for AI agents. So this model is not really built for coding or to be a frontier model or anything like that. It's built to be super fast, lightweight, and small, so that it can code out with AI agents. So for example, if you're using Hermes or OpenClaw or Pi, DeepSeek V4 Flash is actually really good. And the other interesting thing about this as well is you can actually get access to DeepSeek V4 Flash for free via ODE. So if you go to OpenCode, we've got it set up inside our agentic operating system. We can code for free with DeepSeek V4 Flash, and they've updated it to the new version, which means that you can get the latest version of DeepSeek V4 Flash for free and you can code and build out with it. Now also what's interesting about that is there's a skill for Hermes agent from News Research, the founder of Hermes agent, that allows you to automatically delegate coding tasks out to open code, which means that you could use DeepSeek V4 as a brain, you could set up the open code skill, and then when you're coding out with Hermes, you can use open code for free to automatically build cool stuff out as well, which is really cool. I've got another tutorial on that.
5 more minutes of transcript below
Try it now — copy, paste, done:
curl -H "x-api-key: pt_demo" \
https://spoken.md/transcripts/1000651996090
Works with Claude, ChatGPT, Cursor, and any agent that makes HTTP calls.
From $0.10 per transcript. No subscription. Credits never expire.
Using your own key:
curl -H "x-api-key: YOUR_KEY" \
https://spoken.md/transcripts/YOUR_EPISODE_ID