Hermes Agent V0.8: підтримка Google AI Studio та безкоштовний Mimo V2 Pro

[music] >> Hi. Welcome to another video. So, Hermes Agent has just received its V0.8.0 release on April 8th, 2026. And I genuinely think this is one of the most meaningful updates it has had so far. If you remember my previous video, I had already talked about using Gemma 4 with Hermes Agent and Open Claw, especially through Ollama for a proper local setup. But now, Hermes Agent V0.8 changes the story in a pretty interesting way. Because it is no longer just about local Ollama usage. They have now added native Google AI Studio support, live model switching, better GPT and Codex tool use behavior, background task notifications, and they have also added a free Mimo V2 Pro path on NUA's portal for auxiliary tasks. So, this update is not just another small patch. It actually changes how practical Hermes is day-to-day. Now, the first feature I want to talk about is background process auto notifications. This is super useful. Previously, if you gave an agent a long-running task like a test suite deployment, big build, or model training run, you had to keep checking if it was done or ask it to poll. Now, Hermes can get notified when that background job finishes and continue from there. That sounds simple, but it is a really big quality of life upgrade because it means the agent can do real multitasking better instead of pretending to. The second big thing is live model switching with the model command, and this is honestly kind of amazing. You can now switch models and providers mid-session, and not just in the CLI, but also across Telegram, Discord, Slack, and the other gateway platforms. So, if you start on one model and realize you want something cheaper, faster, more reasoning-heavy, or more vision-capable, you do not have to kill the whole flow and restart from scratch. That is a great upgrade for real workflows. Then there is the OpenAI side of things. Hermes says it self-optimizes GPT and Codex tool use guidance by benchmarking its own failure modes and patching them. In simple terms, they found places where GPT or Codex would misuse tools, get stuck, or be too sloppy, and then tighten the guidance. So, if you use Hermes with OpenAI models, this release should make it feel more reliable, which is very good because tool reliability matters way more than some random benchmark chart. Now, let's come to the part that I think is the biggest one for most of you watching this channel, and that is the native Google AI Studio provider. This is huge because if you watched my previous Gemma 4 video, then you know I was mainly talking about the local route. Run Gemma 4 in Ollama, point Hermes Agent to your local endpoint, and then use it that way for privacy, offline work, and zero token bills after setup. That local route is still great, by the way. And as I said before, the Gemma 4 lineup is actually very sensible. The E2B and E4B models are for weaker systems and edge devices. The 26B mixture of experts model is probably the sweet spot for most local power users, and the 31B dense model is the best quality option if you actually have the hardware. That still stands. If you have the GPU and want a proper local stack, Gemma 4 through Ollama plus Hermes is still one of the coolest setups you can build right now. But now, there is another path, and that is why this V0.8 update matters. Hermes Agent now has native Google AI Studio support, and they have also tied that into models.dev, so context length detection is handled automatically. That means you no longer have to rely only on the local Ollama route if your hardware is not strong enough. And this matters even more because Google announced Gemma 4 on April 2nd, 2026, as its most capable open model family to date, and Google says the 26B and 31B models are available in AI Studio. On top of that, Google's own Gemini API pricing page, last updated on April 9th, 2026, currently lists Gemma 4 on the free tier, and it also states that Google AI Studio usage is free of charge in supported regions. So, believe it or not, at least as of now, that means Gemma 4 is not just a local model story anymore. It is also a very practical free API story. So, if you do not have enough VRAM for Gemma 4 26B or 31B locally, you can still use the AI Studio route and get access to Gemma 4 through Hermes without immediately paying money. That is pretty amazing for students and for people who just want to test the model properly before deciding whether to build a full local setup. Now, personally, I still like the local route the most if your system can handle it because privacy, offline access, and not worrying about request caps is always nice. But this native AI Studio path makes Hermes much easier to recommend because now you have options. If your laptop or desktop is strong enough, run Gemma 4 locally through Ollama. If not, use the free AI Studio path. And because Hermes now supports switching models live, you can actually move between different providers in a much more flexible way than before. That is exactly the kind of thing I like to see in agent tools. Now, another new thing in this release is free Xiaomi Mimo V2 Pro on NUA's portal, and I think this is also really interesting. According to the release notes, Hermes can now use Mimo V2 Pro on the NUA's free tier for auxiliary tasks like compression, vision, and summarization. They even mention that Hermes will use Mimo V2 Pro for non-vision auxiliary tasks on the NUA's free tier. This is kind of awesome because auxiliary tasks are exactly the sort of stuff that you do not always want to waste your main model budget on. If the agent just needs to compress some context, summarize a long result, or handle some lighter side work, using a free auxiliary model makes a lot of sense. So now, you could have Gemma 4 as your main local or AI Studio model path, and then still benefit from free Mimo for side tasks inside the Hermes ecosystem. That is a pretty smart combo for keeping costs low, and it also shows that Hermes is becoming more intentional about cost-aware workflows instead of just assuming everybody wants to burn expensive tokens for everything. There are also some other very practical upgrades in V0.8. Inactivity-based timeouts now track actual tool activity instead of just wall clock time, which means if the agent is genuinely doing work, it does not get killed for no reason. That is great. They also added approval buttons for dangerous commands in Slack and Telegram, which is good if you use Hermes through messaging platforms. There is centralized logging now with Hermes logs, structured logs in the Hermes folder, and config validation that catches malformed YAML early. That should reduce a lot of those annoying silent failures and weird config issues. And they also added MCP OAuth 2.1 support plus malware scanning for MCP extension packages, which is honestly quite important as MCP becomes more and more common. So, overall, what is my take on this release? I think Hermes Agent V0.8, 2026, is one of those updates that makes the project feel much more mature. It is not just one flashy feature. It is a whole set of changes that make Hermes more usable in real life. Background tasks are better. Model switching is better. GPT and Codex tool use is better. Google AI Studio support is now native. Free Mimo is available for auxiliary work. Timeouts are smarter. Logs are better. MCP auth is better. And all of that together makes Hermes a much easier recommendation than before. If you ask me, the biggest story here is flexibility. In my previous video, Gemma 4 with Hermes was mainly a local-first story. Now, after this update, it becomes both a local model story and a free API story. You can still go full Ollama and keep everything on your own machine, which is great. But if you do not have the hardware, you can use the AI Studio route for Gemma 4. And if you want free auxiliary help for side tasks, there is Mimo V2 Pro on NUA's portal, too. That is a really strong combination for an open-source agent. So, yes, I think Hermes Agent is on a really good trajectory right now, and this V0.8 release makes it much more compelling for both the local model crowd and the free API crowd. Overall, it's pretty cool. Anyway, let me know your thoughts in the comments. >> [music] >> If you like this video, consider donating through the Super Thanks option or becoming a member by clicking the [music] join button. Also, give this video a thumbs up and subscribe to my channel. I'll see you in the next one. Until then, bye. >> [music]