Chafik Moalem

[Chafik Moalem @ChafikMoalem]

@HermesWatcher @Teknium yup ... I just blinked today and I am more than 4000 commits behind 😅

[Chafik Moalem @ChafikMoalem]

@sudoingX oh yeah? I just dusted off my 2020 MSI WS66 (64GB RAM) with mobile RTX 5000 (16GB VRAM) ... dropped Ubuntu 26.04 on it ... and handed it over via ssh to my Hermes Desktop. A while later : ollama serving qwen3 35b moe with 55tok/sec ... full context. Hermes picked the model too!

[Chafik Moalem @ChafikMoalem]

@HermesWatcher yep, thanks! I have noticed: Hermes can dig very deep! I just tell him "here is a new linux installation rechable via ssh. complete the setup, inspect the machine config and propose hardening. " And it DIGS 😝

[Chafik Moalem @ChafikMoalem]

@HermesWatcher hhh ... I should ask Hermes 😉

[Chafik Moalem @ChafikMoalem]

@HermesWatcher Q: the MEMORY.md (and rules etc) is local to the profile. When I hand over a session to "somewhere else" possibly accross gateways: how does the MEMORY get maintained?

[Chafik Moalem @ChafikMoalem]

@stevibe you need Hermes Desktop bot mode on top of Hermes Agent ... a lot has happened in August!

[Chafik Moalem @ChafikMoalem]

@tonysimons_ Since Nous/Hermes (unlike grokbot) is not locking me into one ecosystem and even does model fall back (in local Hermes) and smart model routing (on Nous hosted models) ... it is for me a NO QUESTION!

[Chafik Moalem @ChafikMoalem]

@richardhutton You need to subscribe with a platform which does smart model routing. Vendor lock is a trap!
The Harness you are using is more important than "which frontier model"!
Checkout e.g. nousresearch (and their Hermes Agent Harness).
Goodbye vendor/model lock!

[Chafik Moalem @ChafikMoalem]

@Lonely__MH the 3090 supports nvfp4 ???

[Chafik Moalem @ChafikMoalem]

@luke_judges Different topic: Hermes Desktop has become a command and control center for bots/bot teams accross multiple gateways ... and your iOS Harness super app is in the make (preview available checkout the Hermes release watcher here on X)

[Chafik Moalem @ChafikMoalem]

@sudoingX Q4_K_M, q8_0 KV uncensored orcarouter with MTP activated (again) in the manifest: 40-50ish token/s with maxed out VRAM on my 3090 with context 114688. With MTP and variying context fill the number is not constant btw.

[Chafik Moalem @ChafikMoalem]

@BkashJosi hmmm ... I just instructed Hermes to inspect my new/fresh ubuntu node (for one more ollama). Maybe its not too late to change it to omarchy 😈 (I will ask Hermes 😁)

[Chafik Moalem @ChafikMoalem]

@HermesWatcher graaah ... I needed that 2 months ago ... had to tweak a qwen3.6 27b to run on my 3090 with a good context lenghth via ollama ... by hand! (I was "new" )
But NOW I gave my Hermes a task: check out my new machine which I made you access via SSH and tell me what to do with it! 😆

[Chafik Moalem @ChafikMoalem]

@RansomWrangler @sudoingX oops ... so its not only me!? 😅

[Chafik Moalem @ChafikMoalem]

@sudoingX for now: Yes!

[Chafik Moalem @ChafikMoalem]

@MichaelGannotti well said!

[Chafik Moalem @ChafikMoalem]

@HermesWatcher yep ... I just noticed today when I clicked on capabilities in the sidebar. keewwwlll!

[Chafik Moalem @ChafikMoalem]

@alexocheema I want to run MORE models locally! 😂

[Chafik Moalem @ChafikMoalem]

@MetaMan Having debt is part of "being rich" game!