Disclosure: As an Amazon Associate I earn from qualifying purchases. This site contains affiliate links.
Hermes Agent on My GX10, Over Telegram
Local LLMs on NVIDIA Spark / ASUS GX10
I use Hermes Agent from Nous Research. Two ways. On my own box, and on Nous Portal.
Here's what that looks like day to day, with receipts.
What Hermes Agent is
Hermes Agent is an agent from Nous Research. It ships often. On Nov 2, 2026 v0.22.6 landed. About 776 merged PRs rolled into stable. New in that window: a Hermes webapp from the gateway, a /model reset that keeps your history, and per skill write locks.
That's a lot of movement. I like tools that move.
How I run Hermes over Telegram on my GX10
My setup is short. Hermes runs on my ASUS Ascent GX10. I control it over Telegram.

I said it plainly on X: "Using Telegram to control Hermes on my Asus Ascent GX10. That's the whole setup." Here's the post: x.com/wikiwayne/status/2117241565600997541.
Why Telegram? It's already on my phone. I send a message, the agent works on the GX10, and I read the answer wherever I am. No laptop needed.
The GX10 side is in The AI Hardware I Own: ASUS Ascent GX10, a Mac mini and a MacBook Pro.
Running Hermes through Nous Portal
I also run Hermes on Nous Portal. That's how a lot of my model tests reach hosted models.
My bead tracker gauntlet is a good example. Three models went through Nous via Hermes for Part 1:
- Muse Spark 1.3 (meta/muse-spark-1.3). Part 1 passed. $0.152.
- Muse Wonder (meta/muse-wonder). Part 1 passed. $0.513.
- Nemotron 3.5 Super VL. Part 1 failed. The recorder never finalized. $0.289.
Muse Spark Open (meta/muse-spark-open) also ran Part 1 through Nous via Hermes and passed. Its Part 2 call failed on the Nous stream twice, so that part went to OpenRouter. I wrote that down too. Routes break. The receipt says which route a result came from.
Nous Portal had limits in that run. It caps requests at 10 MB and takes MP4 only for video. So those models got a smaller re-encode of the same clip.
My game gauntlet runs through Nous Portal too. GPT-6.1 Terra went 7 of 7 on the first try, 41 cents total. Gemini 4 Argon went 7 of 7 first try at $1.3359. More in My Model Rankings and Favorites, From Gauntlet Receipts.
How I like to run my agents
My rule: a strong main model, cheap subagents.
Nous added an Advanced Delegation modes plugin to Hermes Agent. The orchestrator can pick a different model or profile for each subagent, from routes you set up. Cheap model for the grunt work. Strong model for the hard part. That's how I like it.
I also built a router setup on the GX10 with LFM3-1.7B in front of Nanbeige4.3-33B. Same idea, small model first, big model when it counts. That one is in The Local Models I Actually Run on My GX10.
A Hermes bot you can add
One of the Grok Bots on my template page is called Hermes. It's for the Hermes Agent CLI. The page also has Clip Scout, Prompt Lab and Glow Up. That's covered in My Grok Bot Team Runs My Content, Site and Video Work.
What I'd tell you before you start
Updates land fast. Read the release notes before you update.
Pick one control channel and stick with it. Mine is Telegram.
Log the route for every run. Local or Portal. If you can't say where a result came from, you can't trust it.
Read the rest of the stack
Affiliate Disclosure: As an Amazon Associate I earn from qualifying purchases. This site contains affiliate links.
