Hermes Agent Discord + Which AI Models To Run On Each Channel, With Real GoldieBench Scores For Opus 5, GPT-5.6 Sol, Hermes MoA, Kimi K3 And Grok

Hermes Agent Discord lets every person in your server talk to the agent in their own session, and each of those sessions can run a different AI model.

That makes model choice a per-channel, even per-person, decision instead of one global setting.

Below are the GoldieBench scores I read on the live leaderboard on 11 November 2026, what they mean for Discord work, and which model I'd put on which kind of channel.

If you'd like my team to make that call with you, book a free SEO strategy session.

How Hermes Switches Models From Inside Discord

The Hermes Discord docs describe an interactive picker: send /model with no arguments and you get one dropdown for the provider and a second for the model.

Each dropdown lists up to 25 entries, and the picker closes after 120 seconds.

Only allowed users can use it.

The messaging overview adds the part that matters most for a shared server: a /model switch in a gateway chat applies to that session and survives a gateway restart.

• /model <name> changes the model for this session.
• /model <name> --once uses it for a single turn.
• /model <name> --global writes it into config.yaml.
• /model reset drops the override and keeps the conversation.

Because Hermes gives each user in a shared channel their own session by default, one teammate can run a heavyweight model on a long research thread while another stays on a lighter one.

The GoldieBench Scores For A Discord Shortlist

Ranks, averages and "best at" labels come from the GoldieBench leaderboard, which lists 33 models, 50 tasks and 1,310 live demos.

Fusion sits at #1 with 8.59, but it is listed as an OpenRouter multi-model panel, so I've kept it off a list of models you'd pick in a Discord dropdown.

What These Scores Measure, And What They Don't

Per the GoldieBench methodology and the models page, every model gets the same fixed prompts, one shot each, scored 0 to 10 by Julian.

The task categories are games, visuals, sims and pages.

That's a strong read on first-try build quality.

It is not a test of chat replies, Discord etiquette, voice or speed.

Use the gap between 8.27 and 7.08 as a signal of how much quality you give up, not as a promise about any one answer in your server.

Research And Long Threads: Claude Opus 5

Opus 5 is the top single model on the board at 8.27 across all 50 tasks, with 13 golds.

In Discord, research work tends to live in threads, and Hermes gives each thread its own session namespace.

That's the place I'd pin the strongest model: a design agency's strategist tags the bot in #research, it opens a thread, reads back the messages it missed, and works the brief there.

Set it inside the thread with /model, and the rest of the server stays on whatever you chose as default.

Channels That Produce Files: GPT-5.6 Sol

GPT-5.6 Sol scores 8.16 and is marked best at Pages, the category closest to "write me a working HTML page".

Hermes can send generated files straight into Discord as native attachments using inline MEDIA: tags, documents included.

So a #landing-drafts channel for a small shop is a natural fit: staff describe a seasonal page, the agent builds it and attaches the file, and a person reviews it before anything goes live.

The Hard Question: /moa Instead Of A Permanent Switch

Hermes MoA ranks #3 at 8.17 and is best at Visuals.

Hermes exposes it as a one-shot command: the slash command reference says /moa <prompt> runs one prompt through the default Mixture of Agents preset and then restores your session model.

That suits a Discord DM where an accountant wants one careful second opinion on a tricky reconciliation note, without changing the model for the rest of the chat.

Presets and per-slot settings are covered in the Mixture of Agents docs.

Scheduled Digests: Kimi K3 Or Grok

Kimi K3 scores 7.89 over 50 tasks, and its GoldieBench card describes it as tuned for long-horizon agent work.

Grok scores 8.09, though 4 of its 47 tasks are unscored.

Either is a reasonable pin for a cron job that posts a "draft only" end-of-day summary to your Discord home channel, the channel Hermes uses for proactive output.

A plumbing firm might use exactly that: crew voice notes in, a five-bullet digest out for the owner at 5pm.

Pick whichever one your provider setup already supports, and read a week of digests before you trust it.

Busy Help Channels: Where A Lower Score Can Still Fit

Gemini 3.6 Flash sits at 7.08, ranked #18, and is marked best at Pages.

A free-response help channel answering "what time do you open on Sunday?" from a channel prompt doesn't need the top of the leaderboard.

It does need the answers in the prompt to be right.

If the questions get harder, switch that channel up with /model rather than leaving every member on the heaviest option.

Voice Isn't On The Leaderboard

Hermes transcribes Discord voice messages with a separate speech-to-text provider: local faster-whisper with no key, Groq Whisper or OpenAI Whisper.

It can also join a voice channel and talk back, configured through the voice mode settings.

None of that is scored on GoldieBench, so the table above only applies to what the agent thinks and writes, not how it hears or speaks.

Lock Down Who Can Switch Models

Before any model choice matters, decide who may talk to the bot at all. Out of the box Hermes turns away every Discord user; access opens only through DISCORD_ALLOWED_USERS, DISCORD_ALLOWED_ROLES or a pairing code someone requests by DM, which an admin then approves with hermes pairing approve. Those codes lapse after sixty minutes.

Model choice is a slash command, so it falls under Hermes's slash command access control.

List yourself in allow_admin_from, and give everyone else a short user_allowed_commands list; leave model off it if you don't want members changing models.

Members of the AI Profit Boardroom can compare access setups with other members who run Hermes in their own servers.

Remember the wider rule too: nobody reaches the bot until you allow their user ID, role or channel, and authorized users get the agent's full abilities.

For a support-focused version of this setup, see the Hermes Agent for customer support model guide.

My Pick Per Discord Job

The Discord adapter itself didn't change in the v0.23.1 patch, so these picks only depend on the scores, and those move as GoldieBench adds models.

Want the full Agent OS with Hermes already wired up? Join the AI Profit Boardroom.

Also On Our Network

• Agent Operatorsthe eight-step operator setup, from intents to a gateway service
• AI Income Deskfive small-business Discord setups and the risks an owner should weigh
• AI Tool Verdicta nine-criteria scorecard with the strengths, limits and a final rating
• agentos.guidefive copy-paste prompts that make your agent secure and run its own Discord bot

Set a default model, pin Opus 5 on your research threads, and if you want my team to check the plan, book a free strategy session before you open up Hermes Agent Discord to your team.