Hermes Mission Control Dashboard - Ideas, Questions & Feedback

[Daniel-Parke]

👋 Welcome to the Hermes Mission Control discussion page!

We’re using Discussions as a place to suggest ideas and offer feedback for the Hermes Mission Control Dashboard.

If you would like to contribute to the project and submit Pull Requests, reach out and I can add you as a collaborator. No requirements here other than wanting to contribute, I just have it disabled to stop any bot/spam activity.

If you have any questions or just want to chat about some of the Mission Control features, don't hesitate to reach out!

[Ralfhelbig]

what models would you recommend for ollama with patterstage for multi agent automation i havew a ultra i9 285k 128gb ddr5 and rtx 5080 with 16gb vram
and is there a way to easily switch local models from the dashboard

[Daniel-Parke]

So there are a range of models I would recommend trying from the Qwen 3.6, Gemma 4, & Nemotron 3 families.

With 16GB VRAM you can fit all the models I will suggest, and your 128GB RAM should allow for enough offloading to handle 128K context windows. I'm running 24GB cards and I still find using small models for repetitive simple tasks that blow up in context can sometimes be better than the bigger model, due to how much slower token speeds are with large context windows.

Worth playing around with to dial in which models work for your use case of course. Here is the list I would recommend starting with, with the specific model tuning in brackets beside:

MOE Models

These will run much faster than the Dense models, with comparable quality in outputs. Good for agentic tasks:

• Qwen 3.6 35B A3B (UD-Q3_K_XL)
• Gemma 4 26B A4B (it-qat-UD-Q3_K_XL)
• Nemotron 3 Nano 30B A3B (UD-Q3_K_XL)

Dense Models

Much slower but arguably better quality results with more consistency:

• Qwen 3.6 27B (UD-Q3_K_XL)
• Gemma 4 12B (it-qat-UD-Q4_K_XL)
• Gemma 4 E4B (it-UD-Q4_K_XL)
• Gemma 4 E2B (it-qat-UD-Q4_K_XL)

For ALL of these models I am specifically recommending the UD versions over the stock releases where available, as they tend to perform noticeably better. Likewise, I would recommend the QAT versions of the Gemma models where available, as they retain more quality at lower quantisations.

For quantisation, I'd generally recommend Q4_K_XL where it fits comfortably in VRAM, as it offers an excellent balance between quality and inference speed. The main exception is the larger Dense models, where I'd lean towards Q3_K_XL on a 16GB GPU.

The quality difference is relatively small, but the reduction in memory usage gives you much more headroom which increases the chance that the model will fit completely on the GPU, with some room to spare for small context windows. Your large RAM pool will help with this, but at a slower token processing speed, hence why we want to avoid overspill unless we really need to.

There is so much nuance to this that really I would recommend just diving in with one of the models above, and feel free to hit me up with any more questions you might have!🙂