← Back to search

News AI 15/26: Gemma 4 // Seadance 2.0 // Hermes Agent

programmier.bar – der Podcast für App- und Webentwicklung · 2026-04-09 · 28 min
relevance 100 4770 words Episode page ↗ Audio ↗
Show full episode description
Wie hat dir die Folge gefallen? Gut 👍 Schlecht 👎 (Keine Anmeldung erforderlich) In der neuen Folge AI 15/26 begrüßen euch Dennis und Philipp, die die spannendsten Entwicklungen der Woche sortieren. Im Fokus steht das massive Release von Gemma 4, Googles neuestem Open-Source-Streich, der nicht nur multimodal überzeugt, sondern dank der Apache-2.0-Lizenz auch rechtlich für deutlich weniger Kopfzerbrechen in euren Projekten sorgt. Wir schauen uns an, was die verschiedenen Modellgrößen von 2B bis 31B leisten und warum ihr Gemma 4 ab sofort nativ auf eurem Smartphone über die Google AI Edge Gallery testen könnt. Außerdem werfen wir einen Blick auf Seadance 2.0 von ByteDance. Das Videomodell sorgt mit 15-sekündigen Clips inklusive Sound für Aufsehen, musste nach heftiger Kritik aus Hollywood aber in Sachen Copyright ordentlich nachbessern. Ihr könnt es jetzt direkt in CapCut ausprobieren. Für alle, die tiefer in die Welt der Agenten eintauchen wollen, stellt Philipp den Hermes-Agenten von Nous Research vor. Im Gegensatz zu eher konsumentenorientierten Tools richtet sich dieser Python-basierte Agent-Harness primär an Unternehmen und die Forschung. Wir besprechen die Vorteile der engen Ollama -Integration und wie ihr damit hochwertige Traces für das Fine-Tuning eurer eigenen Modelle generiert. Zum Abschluss gibt es wichtige Insights zu den jüngsten Security-Vorfällen: Wir klären auf, warum der Leak von Claude Code doch brisanter war als gedacht und wie eine extrem perfide Social-Engineering-Masche mit gefaketen Slack-Workspaces zum Axios-Leak führte. Schreibt uns! Schickt uns eure Themenwünsche und euer Feedback: [email protected] Folgt uns! Bleibt auf dem Laufenden über zukünftige Folgen und virtuelle Meetups und beteiligt euch an Community-Diskussionen. Bluesky Instagram LinkedIn Meetup YouTube
✨ Episode Outline — click any point to jump to it in the episode
Problem solved
Weekly AI roundup: Gemma 4, ByteDance Seedance 2.0 video, Cloud Code leak and the Axios attack.
Benefits
  • Gemma 4 runs natively on Android and iPhone offline
  • Apache 2.0 license, no legal review needed
  • Strong function calling, structured output and agentic use
  • Seedance 2.0 generates video with synced sound
  • Cloud Code leak reveals reusable harness internals
Use cases
  • Gemma 4 E2B/E4B downloaded (5-8 GB) and run on iPhone 16 Pro in ~2 minutes
  • Gemma free API tier gives up to 1,500 requests per day
  • Gemma 4 E4B matches Gemma 3 27B performance at far smaller size
  • Seedance 2.0 generates clips up to 15 seconds via CapCut or Dreamina
  • Cloud Code JavaScript swapped to OpenAI to run Codex with same prompts
KPIs / results
  • Gemma 4: 26B MoE (4 active) and 31B dense; 128k/256k token context
  • 31B is third best open model, rivaling 20x bigger Kimi/GLM
  • 1,500 Gemma API requests/day; 140 languages supported
  • Seedance 2.0 clips up to 15 seconds
Tools / build
0:00 / 0:00
🌐 This transcript was automatically translated to English from the original.
Hello and welcome to a new programmable AI News episode 15 2026. Today we're talking about Seadance, Hermes Agent and Gemma 4.0. I'm Dennis Becker and I'm joined by Philipp. Hello. Hello Philipp, today without Fabi again you have to say. He was already on vacation last week and I don't know exactly what he's doing today. He's not here today, but I know who has extended their vacation again for Easter or what the reasons are. Exactly. Okay, let's get started. Who would like to start? I think Gemma was probably the biggest one, right? Update that was a bit like Open. Open source model with many capacities. What's new about the model or what do you need to know? So there is now Gemma 4 released last week. I think it was our best release in terms of the partners we worked with. So it's available in every open source tool that you use, from LM Studio to Olama and all other tools on very different API packages. Providers, different clouds, different agents and tools. But perhaps very briefly, Gemma 4 is, so to speak, the successor to Gemma 3. There are four sizes. There are E2B and 4EB again. These were, so to speak, the Gemma 3N models that came in the middle of last year. Same principle here. Multimodal means text, video, audio, image, input, text, output. Can run natively on Android and iPhone. There is also the cool Google AI Edge Gallery app. So you can more or less open the App Store for both right now. How does it know? And search Google AI Edge Gallery and then download the model. Takes about two minutes. Then you can test Gemma 4, 2B or 4B directly on your phone. There is then an MOI variant with 26 billion parameters and 4 active. I would say that it is very good for smaller MacBooks or GPUs. You need something like this, depending on what type of quantization you use, but you can easily run it on, let's say, any modern MacBook. And then there is a 31 billion variant. This is a dense model. That means a little slower relative to the MOI, but a little bit better. The small models have 128,000 token context. The larger models have 256,000 token context. And the very special thing, which I am also very proud of, is that all models are now released under the Apache 2.0 license. This means a common open source license without much research or do I have to talk to my legal team about whether I can use it. Also like the other Gemma models, 140 languages ​​are natively supported. Very, very good. Also received a lot of positive feedback on social media. You can use it in AI Studio in chat. So you can go to Google or AI Studio.com. Then you can use it in the chat. There is also the API variant. That means you currently get up to 1,500 requests per day for Gemma. Of course, depending on how busy it is, that could be a little less. It is available in Android Studio offline. This means that if you program Android apps, you can now use an Android AI Gemma offline without internet. It's available in Light RT, which is a, let's say, C library to run directly natively on the devices. That means if you somehow test out Edge Gallery and think it's cool, you can check out a Light RT. How can I integrate it into my own apps? It is also available in MLX, the special Apple Open Source Machine Learning Library. It's available in Cloud with examples of Vertex Model Garden, Cloud Run, GKE, you name it. It is available for VRLM and on TPUs. So no matter where, I say, there is now Gemma. It's available in Open Claw now too. That means if you bought a Mac Mini back then to run Open Claw and maybe didn't get the smallest one, there's a good chance you can test it now with Gemma and in all the other Agents, I'll say. Cool. How big is the storage size if you download this on the iPhone or Android? Depending on whether you call it 2 or 4B. So you have to say, 2 and 4B are a bit here, so that's why the E in front of them, because they have shared embeddings. I think it's between 5 and 8 gigabytes. But I would have to lie. In any case, my iPhone 16 Pro traded E2B super easily. OK. And can you give a little insight into where we are in terms of benchmarks? So what are tasks that are realistic with this? So the big difference, I would say, to Gemma 3 is primarily Function Calling, Structured Output and Agentec Use. So we also published benchmarks. It's roughly in line with the top open models for its size. So around QN 3.5, I would say, a little better. Here you go, you have to test it yourself. Of course again on Arena, LM Arena very, very good. The 31B model is the third best open model. So it's actually competing with Kimi, K2 and GLM, which are 20 times bigger. But I think you have to test it for yourself. Coding is better, function calling is better. And of course, I think you can use open, small open models for everything, from classification to summarization to other tasks. I think the best thing you can compare is that the E4B model has approximately or better performance than Gemma 3 27B. So we have now moved from 27B to 4B with the performance and then the other models are also better. And base model and fine tune model available. That means if you're currently using Gemma 3 or have used another open source model or open model, then you can just say, hey, okay, I like it fine-tune or I'll test the normal one and then you can use it. Cool. Yes, it's nice that things are continuing in the open world and things are being created there in the same way. Yeah, it's really cool. So I'm always impressed when I run it on my iPhone at how good it is for everyday things. Or the AI ​​Edge Gallery app also now has the ability to run skills. So what we know from Cloud Code and everyone else. You have these agent skills and you can also have them locally on your phone. And I don't think it will be long before you can control your phone locally. So in the sense of saying to Gemma, hey, please do this and that. And then in the background or in the foreground it opens apps and clicks around somewhere and does other things. And that's cool if it's actually available offline and works to some extent. Although that's probably a little more difficult on the iPhone, right? With all the sandboxes or what? Yes, I don't know much about that now. But I mean, Apple and Google are now, I would say, working more closely together. Who knows what will come next. Yes. That's right. You just tried to cook a broken finger into the camera or what did you do? I taught tea about my hand, then cook it on Friday evening. Oh God. Yes, I don't think it's that good. But there are now voice inputs and then I just talk to my computer and it then programs for me. OK. Then get well soon. Well, then we're talking about something that wasn't exactly new this week, but at least as far as availability is concerned, there was a bit of time. Namely via ByteDance’s video model. ByteDance, as a reminder, is the TikTok provider, or at least here in the Western world, TikTok as a big product. And exactly, they have released a video model with C-Dance 2, which is similar to the other big ones we discussed with, I don't know, Video 3.1 or Kling 3.0, I think it is currently at Kling. So a model that not only spits out video, but also provides the whole thing with sound. Clips of up to 15 seconds can be generated. And it actually tops many of the benchmarks that are out there and is at the forefront. Now the whole thing was published for the first time in February, I think, and then it faced very strong headwind, especially from Hollywood at the time, because copyright was simply not an issue at all. So you could use all the well-known actors that are out there and also IPs like, I don't know, Marvel and whatever, you could simply create new clips that, in hindsight, I've looked at a few of these examples like yours in the research, they're pretty awesome. So you don't realize that they look like the original actors, the facial expressions and stuff like that are very impressive. And ByteDance actually gave in and said, okay, then we have to fine-tune it a bit so that it isn't available everywhere. And that's why this rollout was a bit, yes, not so clear, so now it's here. There are somehow also different articles in which countries it is now available and I have no idea. But if I think that, if it's not a phishing site where I ended up now, I don't know, CapCut is such a video cutting tool. ByteDance has also been around for a long time. I even used it too. Very popular for mobile video editing. Exactly, yes. Yes, everything you want to do for social media is actually a very, very cool tool that you can use. And now there is something else called Dreamina. I'm not entirely sure what this product name means, because it's also sometimes called Dreamina Seedance, so to speak. But be that as it may, if you want to try it out, you can also create videos with Seedance 2.0 via CapCut or Dreamina.capcut.com. Do you think it's really fine-tuned or six layers of classification before and after that ensure no copyright material is used? I think it's the latter, right? I think it would also be interesting to know whether it is still the normal model in, say, China Mainland or in other regions. Above all, I have seen a lot of anime, or where people have taken manga and then made a film of it within a few hours. Yes, I'm excited to see where it goes. No, but again it's just worth it if you want to see state of the art, something like that is possible now, what's the quality like there. You might as well just watch a few of these example videos. This is a very impressive model. And then I just remembered, I have two more additions to your last news episode. I heard it yesterday or the day before yesterday, I don't know. On the one hand, to Jan regarding Cloud Code and the source code and why it is news at all and is there anything in it at all. And I think about Jan's sentences, I also looked at it, I'll say a little more precisely. And Cloud Code is really just a rapper around the API. So it's all the prompts, all the processes, how tools are defined, how the system prompt is assembled, how skills are loaded, it's all part of the cloud code, of the source code. That means there are also people who have taken that and not only rewritten it in Rust, but also simply took the JavaScript code and replaced the model provider, i.e. the place where the LLM call is executed, with OpenAI and were then able to easily use codecs within Cloud Code with the same system prompts, with the same features more or less. So it's really big news. Especially because as a, let's say, framework provider or body you can really look, okay, how did they manage to get the cloud to sometimes do this or that or there was also this auto mode that came into Cloud Code at some point with I no longer have to do Skip Dangerously Permission and that's also part of the code, I'll say and they use Haiku to make requests like that, so to speak. So it was a very big leak and not easy, I have an API here called Cloud Code and it's running. The same thing is that Cloud Code is used as the code base for the Agent SDK. So it's more or less just a hook in there and it wasn't available on the client before, let's say. So of course, you somehow had the types, but the compiled binary code in BAN, you couldn't do anything with it. Sure, you could have gotten the prompts out somehow, you had done that before, but how everything was put together was a very big topic. And the other thing is, there was this Axios leak and I haven't heard the episode change yet, but I don't know if it was discussed. We now know how they got the data. Did you have that? I don't think so, no. Okay, there's now a postmortem on GitHub, I might as well check it out. Namely, they went to a contributor, they approached him as a founder, with a complete fake company, which they cloned because it looks so similar to other startups. Then they added a real Slack channel to them, with a real workspace that was branded, that had CICD channels where people were chatting, which all looked very real. They had channels where they shared LinkedIn posts or Twitter posts that were interesting for the company. And they even created fake users, fake profiles more or less in Slack, which then wrote to the person. Then they had a meeting with him, not in Slack, but in Microsoft Teams. And the meeting had several people more or less, but it didn't work when he opened the link. And then they sent him a new link that downloaded something. And that was the RAT, the ransomware that he then installed under the assumption, hey, I'm going to go into this Microsoft Teams meeting. And then his credentials were gone, so to speak. So it's not like I somehow leaked my token on GitHub or I clicked on some phishing email that I received. It was really company faked, founder faked, new Slack collaboration workspace opened, different channels, different fake people and people who worked in there for several days. Yes, awesome. Then a calendar invite, so to speak, via Microsoft Teams so that we can talk and discuss something. D, more or less faked the site and then made a download link to a bash file. And when I think about it, I've worked with a lot of companies and startups. That's, so it's a little bit of a bigger vector that you can imagine, especially with projects that are so popular. Yes, absolutely. So I think I don't even know whether we can, in the end, protect ourselves from this. So it's just this, this social engineering, where you somehow get into people's minds and then try and install something somewhere for weeks. I mean, there has to be people who are super suspicious of people and don't trust anything anymore and check everything and have no idea what. I think it kind of becomes a part of it, yeah. But I think that, two or three years ago, rebuilding an entire startup company with a product and a Slack channel would have been a completely different topic than it is today. So just the fact that they proactively shared LinkedIn and X articles that were legit is such an issue for me. Okay, this could also be the programmable channel that we have somehow, something like that. Yes, no, absolutely, that's right. That's right. Wasn't it even recently in the programmable channel extension? That sounds like what I wrote about Prompt Inject. Yes, ignore the next posts, Philipp. I'll do it anyway. But yes, plus the two topics. This is definitely very, very interesting. what's happening now and what you have to pay attention to and maybe the internal, I don't know, somehow more tooling we need. Yes, that could be something that is monitored even more automatically or something like that. Or did you then again... Yes, so many have written afterwards, so hey, you can adjust your NPN config so that you only see updates, so that you see updates that are seven days older. Yes. Then someone else says, but okay, what if you really have a CVE or some kind of leak, like with the React Server components for example, then you have the fix seven days later and then you have the other attack vector that the updates are not installed. So it's not an easy topic. Yes. Good, then you brought the Hermes agent with you. Exactly, so we've had Open Claw, Clotbot, Moldbot, whatever. Yes. And Hermes Agent is an alternative from Newest Research. Maybe some of you have already heard the name. Newest Research has always been known for the Hermes Open LLM models. So, especially based on Lama, they had very, very good Open LLM models back then and they now have Hermes or, above all, it is a bit older, I would say, or is currently being built. Hermes Agent Released, which is an MIT-Licensed Agent Harness, very similar to Open Claw, means that you also have Telegram, WhatsApp, various channels as input, you have a CLI with which you can chat, you also have this single gateway principle, which means that I can send messages to my agents from different channels, you have skill support, you have lots and lots of other features that you know with Cron and other topics. What's special, I'll say now, is that it's written in Python and it's more or closer to research oriented or companies oriented than Open Claw. Open Claw is very B2C oriented, I as an individual user use it and then do certain things with it and what they paid attention to, if I understand it correctly, it is Research Ready, which means you have more programmable interfaces, you have batch generation, you can incorporate it into your reinforcement loop, all the interactions you do with Hermes Agent are saved in a format that I can also use to then train smaller, other open source models, so to speak, that is, everything is designed so that I can use the agent in my training of my own models, or in generating data for Evals, for fine tuning and all of which, I would say, fit more or less into the Python ecosystem. Interestingly, there is also a migration from Open Claw, which means that if you are currently using Open Claw, then you can relatively easily take a look at Hermes Agent and they have, because they are, I would say, Open Lab, more or less a strong focus on the Open models, that is, they work with all the, I would say, Chinese Open Model providers. There is support for Open AI and Cloud and Gemini, but also for Gemma now and is definitely interesting, so you have a similar onboarding principle. I have a CLI that I install and then I can just go through an onboarding floor where I can select my providers and very cool, especially because they also have a relatively tight integration with Olamo. I think this is somehow a command that I enter in order to be able to use Olamo as a bag end with the Hermes agent with the new Gemma model now and then I have this full-plan agent harness, which I can then link to Telegram or other things very quickly and that I have these traces for every interaction with memory, which I can then really use for training. OK. And then just the same features, i.e. with skills, with memories. I say it myself that they create skills more on the Hermes Agent, because as users they create skills and they are also tested or executed themselves in the background. So definitely, if you're Open Glove fans, please take a look. Maybe you'll learn something new, maybe it'll fit a little better and especially if you're somehow in, I'll say, more company space where you might be able to use an agent at some point, then it's a huge advantage if I can now start to create a little bit of my own data about how my users, how my, I'll say, developers, how something else interacts with this agent. Yes. Somehow I always feel like there's such a stark difference. Whether it's still in the private space at the moment, which you have in terms of all access and other things, or if you think about it at the enterprise level, yes, there are always a lot more implications, where is the whole thing and what are the accesses and how is the user management and who can then access the data and how is it all shared and so on. That somehow opens up a completely different space of thoughts that you somehow have. Yes, now, why I find it particularly interesting, there is currently such a growing focus on it, everyone uses agents for all sorts of things and, let me say, two years ago we all created fine-tuning datasets, synthetically with LLMs and everyone shared these fine-tuning datasets. There was ShareGPT, which at that time was real chats from ChatGPT, which users shared and stuff like that and everything goes in the direction of hey, okay, with these agents, we need all these agent sessions where you have several user inputs, where you have the whole tool calling loop, so that I have that as data to evaluate, to get real prompts, to then improve models and that's where everything goes. Now I don't know whether OpenClar also has a feature where I can export it directly or whether everything is generally broken locally so that I can use it directly or something like that, but I found that very, very cool for everyone, I would say, model builders, people who use fine-tuning models. Especially now with Gemma you have a relatively good, smaller open model, which is very good at function calling. If I now somehow use Hermes Agent in my team or just privately, then I can probably do it better very easily or then just improve the data that I have collected with other models or in the sense of hey, I somehow have a trajectory here where I say, hey, please, I don't know, check my Gmail calendar, my emails and create a daily flow or plan my new meetings, find time with Dennis and if I have it, it might not be perfect because the model somehow makes tool calling errors or something like that, I can adapt them or change them synthetically so that I have the right flow and then train models on them again or then use reinforcement learning to verify them and I think that's the biggest, I'll say, difference between them and I also say myself, of course, they're better and it's a bit leaner and but then, I'll say, it all depends on personal preference. OK. Do you have anything going on like that at the moment? Nope. So not personally now. OK. I don't know, I still use Poke. Yes? Yes. Cool. OK. But also more like, I don't know, I don't have as many things that I have to do in the background now where I need a notification, I have to say. Yes, what were they? Pika. Pika, I think we talked about that at some point. They made, above all, short video clips with ready-made scenes and stuff like that. They also released an agent as an app the week before last or something like that. I just think so, I can imagine it as a very stripped down version of one of these things, but where you can also give access to a few services and it's just somehow rapped in a nice interface and it just talks about your AI image, so to speak. This means that you also take a snap, a selfie of yourself and then a 3D avatar is created that looks like it and it also has a certain business connection. Says, hey, if you kind of want to post stuff on LinkedIn or I don't know, yeah, just try to leverage it a little bit. but I still didn't like it that much. what I saw, so I have, Latent Space did a very good podcast with the Tremern Startup. Tremer was directly acquired by Meta within a month of their launch. This is an Agent OS startup, you can listen to it. They also talked a lot about it with these, let me say, proactive background agents, what they use them for and I found interesting use cases, especially if you work a lot or are busy, that you have topics like, I don't know, what can I do with my children on the weekend, so they just said, okay, you get some kind of report every week, based on what they've done in the past and directly suggestions or for everything Others, I don't know, I just ask Gemini somehow or if I want to book a flight, I just book it myself and privately I don't have that many emails and I'll say, social media connections aren't always good and somehow you don't want them to be with you, so if someone writes me a message, then I answer them. Then I don't somehow have a bot that answers, that's a bit, that is, a lot of opinion is simply more useful in the professional environment than somehow in private. I hope that all smartphone devices will now receive upgrades more quickly and then there will be more usage. Wait. Yes, that's a good question. I mean, I somehow spent two or three hours yesterday looking for a new house for a ski vacation next year. Yes. Do you buy directly, right? Two weeks of skiing vacation, then a short one. Yes, yes, exactly. Who is renting? No, for rent of course. Good. Then we have it. Yes. Thank you. Then have a good trip. You'll soon be traveling on planes again for work. Yes. And then hopefully we'll see each other again in two weeks. Otherwise write, yes. Yes. I think Google Cloud Next will be in two weeks. I don't know if the programmable one will be there, but I will definitely be there. You will definitely be there and Jan will also be there. Yes, perfect. Very good. Cool. Then we'll see how we can maybe fit a recording in there or not. Very good. See you soon. Take care. Ciao.