NEW DeepSeek V4 Flash Update is INSANE!

A brand new version of DeepSeek, DeepSeek 4 Flash, just went live and it's built for AI agents. You can see the announcement just happened a few hours ago today. DeepSeek just dropped a major upgrade to V4 Flash and it makes your agent seriously more powerful. So DeepSeek say the new benchmark scores far surpass their previous top preview model from the small, fast and cheap tier.

I'll show you exactly what I've built. We've actually built over 50 things of it already today. So we've tested it relentlessly and that means your agents get to think across a million tokens of context. They run longer coding loops and finish more work before you touch the keyboard with AI agents.

It's tuned exactly the way modern agents work, so it slots straight into the tools you already use. And today I'll show you what changed, what builders are saying about it, how I've wired it into my own agent operating system in minutes and I'm running it through my actual benchmark. So stick with me because the last part is where it actually gets interesting. Let's get into it.

So you can see the new announcement right here and how it performs. So it's a massive upgrade from Flash preview. You might actually wonder like what has changed here. So previously V4 Flash was in preview mode and now the official API is actually live.

So you can see how it performs. Big, big step up from 61.8 to 82.7, 39.4 to 54.2 on NL2 repo, Cybergym 38.7 to 76.7. Huge upgrades. DeepSWE 7.3 to 54.4.

By the way, if you're wondering what does that even mean in reality, how does that compare to other models? So they've actually compared it to Opus 4.8. Obviously Opus 5 just dropped last week, which is obviously much better and kind of like Fable 5 level. But you can already see like it's kind of matching with Opus 4.8.

So I'll show you how it works in a second, what we built with it. Let's get straight into it. So we were testing out relentlessly on GoldieBench because as always I don't want to just look at the benchmarks. I want to actually test it out.

So we've built our bunch of stuff of it here. Let's have a look at one. So this is like a parachute drop game as you can see. I will say it's not like, this is not Fable 5 level peeps.

Like just so you're 100% clear here. This is not like a frontier level model and this isn't like at the same level, something like Fable 5. At the same time it's very very fast. So I mean we literally built out 50 builds like this in the space of a few hours and it can create some pretty cool stuff.

Let's take a look at another one. Here's another example. So you can see this game right here. It's fun to play.

It's interesting. By the way you might be wondering like why are you building like games with this? Why are you not just using for example business automation? So showing that sort of stuff.

Mainly because it's like a quick way to get a feel for the logic, for the planning, for the UI, for the design, for how fast the model can run and how it can handle complex prompts. So that's the reason that we build this sort of stuff. Obviously in reality you're going to be using it for actual useful automations, landing pages, websites, agents, etc. And that's really what it's designed for.

So we tested out on the new API. Pretty good, pretty fast. Most of it rendered pretty cleanly as well. Again not Fable 5 level but it can create some cool stuff.

And so this is the same model but just a much better agent. And this is a path worth understanding. Normally when a model gets better, it gets bigger, more parameters, more hardware, high price, etc. But DeepSeq were pretty clear about this.

V4 Flash uses the exact same model architecture and size as a preview. The upgrade is in the training, not the size. So it's like the same speed, same tier, just a much sharper agent. So it's much better at agentic stuff.

So for example if you're plugging this into your AI agents like Hermes, that is what it's ideal for. That is what it's designed for. It's not designed to be like a flagship coding model. By the way V4 Pro, that release is still coming.

I'm excited to see what happens there. That could be Fable 5 level when DeepSeq V4 Pro actually gets released. Now we've actually already added it to our agent operating system. I'll show you exactly how you can use it.

So we've added it in two different ways and we also tested it on GoldieBench. So number one is the coding tab. So we've actually created this DeepSeq coder as you can see right here. And if you want to get this system, it's inside the link in the comments description or go to the AIProfitBorne.com.

So we've plugged in DeepSeq V4 Flash and then you can code with it over here. You can ask it to build something and then it'll actually appear as a preview over here and you can grab the code for it. Now also we can see all of our previous builds inside this section. So we can see what we've created with DeepSeq so far and the great thing about that is that everything that we create is saved for later.

We can view it later. So let's have a look. You can see all of these builds. We can open up in full screen.

It is capable of creating like 3D games as you can see right here and this is within like a few hours. So don't get me wrong, I didn't get a lot of time to test this but at the same time it looks pretty good. Like you can create some cool stuff as you can see right here. This is like really smooth, really fluid 3D game.

Works pretty nicely, easy to use. There's no bugs in that at all. I've seen much worse outputs from frontier models in the past. So we've got the DeepSeq coder over there so that we can code with it directly and then save everything to our workspace.

But then also I've been testing it inside Hermes right. So if we go to DeepSeq, we've got a separate profile for Hermes agent. This is usually what I do. So I'll create separate Hermes profiles for each model that drops so that we can just plug them into our system and we can test them and use them whenever we want.

So you can see for example we can use this directly with Hermes agent. If you're wondering how to change the model inside your Hermes agent so you can use DeepSeq, you can just go to your terminal like so, type in Hermes model and you can change it right there. And that's really the way it's designed to be used. Something that's really important to note here is that you want to make sure you're actually running the new one.

So if you've already got access to for example like OpenCode with DeepSeq v4, just make sure that you're not using the preview model and that you've actually switched to the latest version. Easiest way is just to get the official API from DeepSeq. So it's the same size, same name, just a better agent that performs better on agentic benchmarks. Now just to be very clear here, you can see that today's upgrade only applies to the DeepSeq v4 Flash API right.

So the DeepSeq v4 Pro API and app and web models remain unchanged for now. So that means if you go to DeepSeq.com and you're trying to use the latest models, it is not the latest version. This is only via the API and the official release of DeepSeq v4 Pro is coming as well. So I'm excited to test out.

I think that'll be a real step up if today's benchmarks are anything to go by for DeepSeq v4 Flash and this is DeepSeq v4 Flash 0.7.3.1. So it's just the date of today and as you can see they've fully confirmed it's the exact same model architecture and size as a preview version. Just performs way way better. Also the open weights should be released pretty soon as well as you can see from this update by Klein.

Other thing to note is like it's way cheaper than for example Fable 5 or something like that. So if you compare Sonic 5 versus Fable 5 versus GPT 5.6 Soul, look at the price and difference. It's absolutely massive. Also if we look at the leap on the artificial intelligence analysis index, this was the old version and it has leapfrogged a lot of models to move over here.

So it's still not frontier but it's not far. It's kind of in the middle, just a lot cheaper. 60% lower than GPT 5.6. Luna Max which is a model with comparable intelligence and it's within one point of GLM 5.2 which was still a pretty good model, still an impressive model and seven points behind the open weights frontier Kimi K3.

It's also in line with a Gemini 3.6 Flash and one point behind Muse Spark 1.1 with a million token context window. Also once the weights are released which are coming soon, this will be the second highest open weight score. By the way this is a really useful tweet by artificial analysis just giving you a full breakdown of the scores here. So once weights are released this will be the second highest open weight score behind Kimi K3 ahead of GLM 5.2.

So very good at agentic performance and also token usage falls 12% against the predecessor. So it's more token efficient whilst achieving better intelligence. So they've done some really interesting stuff here and it improves over the previous preview version on every evaluation in the intelligence index. Again test out for yourself don't believe all the benchmarks but from what I've built so far it looks pretty good.

If you want to see everything that we've built it's directly on GoldieBench. Not all of these builds work so for example this one it seemed to totally get stuck but overall pretty good model pretty useful great for agentic work excited to test it with Hermes Agent as well and we've already built it into the AgentOS. The other thing I would say here is like don't go around chasing the models you want to have a good system like this where new models can just be plugged in and swapped out as soon as they drop. So this only dropped a few hours ago we've already plugged it into the AgentOS super simple it's already got access to our memory system it already has the DeepSeq coder built in now and we've got all of our builds saved over here plus it can work alongside all of our other agents like Claude, Hermes, Antigravity and everything else.

So if you want to get our full system feel free to get it inside the AIProfitBoardroom.com link in the comments description and you can get a full AgentOS there. So if you go to the classroom go to new daily updates you can find the AgentOS with a video tutorial the zip file to install and then we add new daily tutorials based on what's actually useful and just dropped. You can also post questions inside the community and I'll personally answer them and then inside the calendar you can jump on weekly coaching calls, share your screen, ask questions, meet other members. Inside the map you can meet people in your local area who are building with AI agents like Hermes, DeepSeq and other tools and this is all inside the AIProfitBoardroom.com.

You might also say like is an AgentOS or is this stuff sort of technical? I'm non-technical and I can use it and loads of people getting amazing results with this stuff. So we've got over 205 pages of wins and testimonials and reviews from AIProfitBoardroom members and people absolutely crushing it with this stuff. So if they can do it and they're non-technical and I'm non-technical too you can use AI and get the most out of it too.

Hope to see you inside there. Cheers for watching. Bye bye.