← Back to search
DeepSeek V4.1 Flash + Hermes Agent Is INSANE!
AI News Today | Julian Goldie Podcast · 2026-10-19 · 8 min
Show full episode description
DeepSeek V4.1 Flash + Hermes Agent Is INSANE!
✨ Episode Outline — click any point to jump to it in the episode
Problem solved
Frontier-quality agentic coding with
Hermes Agent has been slow and expensive;
DeepSeek V4.1 Flash 1017 makes it fast and cheap.
Benefits
- Frontier-level coding at Flash speed — builds finish in minutes, not tens of minutes
- Roughly 55x cheaper than Opus for identical token counts
- No more Hermes Agent timeouts on big builds caused by slow frontier models
- Swap model 'brains' per task inside Hermes — free, cheap, or frontier
- Better post-training only; same architecture as previous V4.1 Flash
Use cases
- Ran 'Reverse Engineer Anything' on an AEO/SEO client portal, then rebuilt the whole app on localhost in ~5 minutes (10:38 → 10:44)
- Built a 3.js diving game and a 3D racer game with Hermes Agent + DeepSeek 1017
- Generated four websites from a single prompt
- Three Hermes Agent builds cost $0.38 total vs an equivalent $21 on Claude Opus API rates
- Side-by-side: same DeepSeek model in Claude Code vs Hermes Agent gave different designs — Hermes output had animated backgrounds
KPIs / results
- $0.38 for three builds vs $21 equivalent on Opus
- Average build time ~14 minutes
- Client portal rebuilt in 5 minutes
- Outperforms GPT 5.6 on most benchmarks; near Opus 5 on coding benchmarks
Tools / build
- Hermes Agent with DeepSeek V4.1 Flash 1017 profile
- Reverse Engineer Anything (GitHub repo) plugged into Hermes
- Agentic operating system with switchable model profiles
- OpenCode and DeepSeek Harness/DeepSeek Coder integrations
- Claude Code with DeepSeek 1017 for comparison
Today we are going to be testing out DeepSeek V4.1 Flash 1017, the new update that just dropped from DeepSeek, along with Hermes Agent. It is designed to be a genetic model, so you can see for example already plugged it into Hermes Agent inside our operating system right here. And so we've got Hermes Agent working with DeepSeek 1017, and you can see some examples of how we've already used it here. So for example, I said, you know, use reverse engineer anything, which is a new trend in GitHub repo that's just dropped to open source on our client portal here, right? Because I wanted to reverse engineer our client portal where we basically have AEO services and SEO services, as you can see. And then we plugged that into Hermes Agent, and it basically reversed engineer exactly how it worked, but not just that. I said, can you create this on a local host for me? And it actually recreated the whole process based on what it understood. So you can see this hosted on a local host, and it just recreated the whole page, basically rebuilt the whole app, even has the whole setup here. So, you know, people can select between the different packages that they want. And that's literally what done in about 10 minutes is unreal. It's really, really powerful. I mean, for example, I think we dropped them. This is the other thing about DeepSeek V4.1 Flash, the new version is like it's fast, but it works at frontier level, which has never really been seen before. So you can see we messaged it at 10.38, and then at 10.44, five minutes later, it already created this whole app, which is insane. So the other thing to note here as well is with 10.17, it's basically performing at frontier level. It's outperforming GPT 5.6 Soul on most of the benchmarks. It's up there on coding benchmarks with Opus 5, and this is an open source model that works a lot faster. So it's a much faster model to use. And it's also massively improved versus the previous version of DeepSeek V4.1 Flash that dropped in September last month. The other thing to note here is that it's a lot cheaper to use in other models. And on top of that, this hasn't changed in terms of architecture. The only difference that's changed with the new version of DeepSeek V4.1 Flash versus the old version is that it just got better post-training. Now, you might be wondering, OK, well, what does that look like in reality? Well, I've already shown you one example of the portal page we created. But also, if we go over to our builds here, I can show you some examples of what we created. So if we open this up, this was kind of like a fun diving game, as you can see right here. And the thing that I'll say is DeepSeek V4.1 Flash 10.17 is much more visual. It's capable of coding with 3.js. It can create some nice designs. It's not as good as Opus 5.5, I'll be 100% honest. But for a quick creation like this, it's pretty awesome, right? You can create some awesome stuff. Let's have a look at another example. So here's another example. And we actually compared this side by side. So if we go over to Claude Code versus Hermes Agent, the other thing that I'll say here, and it's pretty interesting, is the creations here are quite interesting. So what I'm trying to tell you here is depending on the agent harness, things will change dramatically in terms of what you create and the outputs you get. So for example, this is Claude Code with DeepSeek 10.17, right? And this is Hermes Agent with DeepSeek 10.17. Now, if we have a look side by side, the designs are subtle but different. So you can see the moving background right here, and it just feels a lot nicer to use. If we scroll down to this page, it feels a lot more flat, and it doesn't feel quite as fun to use, right? And so it's got these nice image animations on the background. It looks pretty cool. It's just got a few nicer touches inside Hermes Agent where you can use it nicely as an agentic harness. Now, if you want to test these all out inside one place, what I would recommend is that you can have Hermes Agent, and we've got an agentic operating system, as you can see right here. We give it to members inside the AI Profit Boardroom. Link in the comment description or go to theairproport volume.com. And basically, inside the setup, you can have multiple different profiles. So, for example, if we type 10.17, we can get DeepSeek. If we type in Claude, we can get Opus 5.5, right? And so we can switch between models and profiles really, really quickly. Or if you want to use, for example, Claude, we can use Claude here, and we've got custom-made workflows, all sorts of amazing stuff that just makes it way more fun to use than Claude. And the final thing that I'll say is you can use OpenCode with DeepSeek 10.17, as you can see, or you could switch over to something like DeepSeek Harness, which we've also built into this system. We've got DeepSeek Coder and then DeepSeek Harness, as you can see. And you can basically use DeepSeek 10.17 in multiple different ways, just to see which one gets the best outputs here, which I think is very, very important. So when we compare them side by side, pretty cool. I mean, Hermes can definitely code right now, especially it can code a lot faster. The other thing to note here is like when you're using previous Frontier models with something like Hermes, if you were creating a big build, it would quite often time out with Hermes Agent because the model you were using, if it was Frontier, would be quite slow. Whereas with DeepSeek v4.1 Flash 10.17, it's much faster to build because it's not going to time out and it's a Flash model. And so you can build at Frontier levels, but you can build with Hermes Agent and code actual decent stuff in minutes rather than, you know, like tens of minutes. Now, here's another example. So this is a page that we created right here. It's quite good at creating nice designs, as you can see. So it can build out like four beautiful websites in like one single prompt. And you might be wondering, okay, well, how much did this cost? And here's the crazy thing as well. So for all three of those builds that you just saw with Hermes Agent, it cost us $0.38. But if you use the tokens on Opus, that would cost an equivalent of $21. We actually compared it side by side. So we actually got Claude to measure the differences. But the identical token count on Claude Opus on the listed rates via API will be $21 versus $0.38, which is insane. And the average build took about 14 minutes from all the stuff that we build today, which is pretty wild. Now, why would you use Hermes for this? Well, basically, you've got one agent and then you can quickly switch out the brain. So it's a lot more flexible. Like I was showing you before, you can switch to all these models. You could use a free model if it's just a basic task. And then when you want something more Frontier but cheap, then you can switch over to DeepSeq if you prefer. Also, it's got its own tool belt. So it's really easy to add like connectors to Hermes Agent now. It's got a really good memory and it self-improves all the time. You can talk to it anywhere because you can connect it to Telegram, Discord or whatever you want. It's built for cheap models. So it's built for stuff like DeepSeq. In fact, DeepSeq themselves on their benchmarks and reports recommended using it with agentic harnesses like this. And also, you can just plug it into Claude code later if you prefer. That's what we did at the same time. So that's basically the system, how it works, what we built with it. Let me show you another example of what we created. So if we go to Claude, I actually test it out with open code over here. And then if we go to the workspace and let's open up this 3D racer game that we created with DeepSeq as you can see. And the quality in the graphics, not quite as good as Opus 5.5. Just want to be 100% clear there, right? So don't ask me, is it better than Claude Opus 5.5? I wouldn't say so. But is it cheaper? Is it faster? And does it build comparative stuff? Absolutely. Like, you can get a lot out of this model, right? And it's pretty impressive what it can do so far. So for me personally, like 100%, is it a good model? I mean, DeepSeq V4.1 Flash was already really good. But now that you plug it into Hermes with the 1017 update, it's massively improved. And also, this is outperforming stuff like GPT 5.6. Now, you also see, if you look at the benchmarks from DeepSeq, that they didn't compare it against GPT 6 Astra when I look at the tests. So it seemed to be just like GPT 5.6 Sol and Opus 5, which are previous generation models, really. You know, like Opus 5.5 is out now instead of Opus 5. And GPT 6 is out now instead of GPT 5.6. But still, the gap between open models, the gap between DeepSeq, the gap between these frontier models is just getting bigger and bigger. And this is like the first time we've really seen a fast, lighter model work really well at frontier level outputs, right? So if you want to get our full agentic operating system, all of my best trainings on DeepSeq, all of my best trainings on Hermes and everything else, you can get that inside the AI Profit Boardroom. Link in the comment subscription or go to the AIProfitBoardroom.com. This is my AI automation community for learning how to save time, grow and scale with AI automation. Inside the community, you can direct message me and ask me any questions. You can also post inside the community here. I'm always happy to help. In the classroom, you can get access to all of our new trainings, as you can see, and our new daily updates. And then in the calendar, you can jump on weekly coaching calls, ask questions, get help and support in real time. And in the map, you can meet people near you who are building out with AI agents like you. So feel free to get this. Link in the comment subscription or go to the AIProfitBoardroom.com. I'll see you in the next one. Cheers. Bye-bye.