Hermes Agent Telegram + Which AI Models To Run Behind Your Bot, With Real GoldieBench Scores For Opus 5, GPT-5.6 Sol, MiniMax M3, Kimi K3 And Gemini 3.6 Flash

Hermes Agent Telegram runs on whatever model your Hermes profile points at, so the model you pick decides how good every reply on your phone is.

GoldieBench gives a clear signal for that choice: across the five single models on my Telegram shortlist, averages run from 8.27 down to 7.08 out of 10.

Below are the scores I read on the live GoldieBench pages on 11 November 2026, what they do and don't say about chat work, and which model I'd put behind which kind of Telegram use.

If you'd rather my team made the call with you, book a free SEO strategy session.

Where The Model Comes From On Telegram

The Telegram bot doesn't have a model of its own.

It's a front door to the Hermes messaging gateway, which runs your configured model for each session.

That gives you three levels of control, per the docs.

1. Per chat: send /model and Telegram shows an inline keyboard of providers, then models, with Prev and Next buttons.
2. Per channel: channel_overrides sets a model and system prompt per channel ID from one gateway.
3. Global: the default in config.yaml, or /model NAME --global to write a chat's choice through to it.

A chat's /model choice beats the channel default, and it now survives gateway restarts.

The overview's channel_overrides example uses Discord IDs, so test it on a Telegram group ID before you rely on it.

The GoldieBench Scores For A Telegram Shortlist

Ranks come from the GoldieBench leaderboard, which lists 33 models, 50 tasks and 1,310 live demos.

Fusion (8.59) and Hermes MoA (8.17) sit higher, but they're multi-model panels rather than one model you'd set as a chat default, so I've left them off this list.

What A Build Benchmark Says About Chat

The methodology page is clear about what's measured.

Every model gets the same fixed prompt, once, and must return one self-contained HTML file.

Scores are Julian's own 0 to 10 verdicts, averaged across every task a model was scored on.

So the table measures first-try quality on hard build tasks, not how well a model answers a quick Telegram question.

Read it as a signal for the heavy requests you send from your phone: drafting, summarising, planning.

It says nothing about speech: voice notes are transcribed by a separate engine before the model ever sees them.

Speed isn't scored either, so check each model page's trade-offs for notes like "slow on hard tasks" before you put it behind a chat people expect quick answers from.

Heavy Drafting From Your Phone: Opus 5 Or GPT-5.6 Sol

When you voice-note "turn this into a proper proposal", you want it right first time, because editing on a phone is painful.

Claude Opus 5 has the highest single-model score here: 8.27 across 50 tasks, with 13 golds on its page.

GPT-5.6 Sol sits close behind at 8.16, and its page says it was benched at medium reasoning effort.

The Opus 5 page lists "reasoning-by-default eats token budgets" as a trade-off, which matters on a bot people message all day.

That's an argument for giving the strong model to one channel, not to every chat.

Long Voice Memos And Big Documents: Kimi K3 Or MiniMax M3

A long transcript or a hefty PDF needs room.

Both pages list a 1,048,576-token context.

Kimi K3 scores 7.89 across 50 tasks, and its page says it's wired into the Agent OS as the kimi-k3 Hermes profile for long unattended agent runs.

Its page is honest about the cost: slow on hard tasks.

MiniMax M3 scores 7.97 across 47 tasks, and its page says it's designed for high-volume agent workloads where context length matters.

Remember the file side too: Telegram's public Bot API stops at 20 MB, per the Hermes Telegram docs, and a self-hosted Bot API server lifts that to 2 GB.

Quick Replies All Day: Gemini 3.6 Flash

Most Telegram traffic is small: "what's on today?", "reword this", "remind me at four".

Gemini 3.6 Flash scores 7.08 across 50 tasks, and its page tagline says it uses "fewer tokens".

The same page lists it as best for routing the everyday 90% while a flagship handles the hard 10%.

That's the pattern I'd use: Flash as the default, a flagship one /model tap away.

Switching Models Mid-Conversation

You don't need to pick one model forever.

Inside any Telegram chat, /model opens the picker, and the current model and provider sit at the top of that message.

The whole picker edits a single message in place, so your chat doesn't fill up with menus.

Know the exact name already? Type /model NAME and skip the buttons.

Add --once and the switch applies to a single turn, which suits a one-off heavy request in an otherwise light chat.

/model reset drops the override but keeps the conversation, while /new clears both.

That makes the routing table below a starting point you can override with one tap, not a rule.

The AI Profit Boardroom trainings show how members set these defaults per profile inside the Agent OS.

The Routing I'd Set Up

Scheduled jobs can carry their own model pin, which our Hermes cron job model guide covers in full.

Three Small Businesses, Three Model Choices

A print shop owner

1. Sets Gemini 3.6 Flash as the default for his DM.
2. Sends quick questions between jobs: reword a quote, list today's pickups from a notes file.
3. Gets short answers that don't burn a flagship's budget.

A physiotherapy clinic's admin group

1. The clinic gives its admin group a channel_overrides entry pointing at Claude Opus 5.
2. Staff mention the bot to draft a patient letter from bullet points.
3. A clinician reviews and sends every letter; the bot only drafts.

A wedding photographer

1. Records a long voice memo after each client meeting.
2. Switches the chat to Kimi K3 with /model for the summary.
3. Gets a shot list back and keeps the memo under Telegram's 20 MB limit.

Settings That Matter More Than The Model

A top score won't help if the bot can't see the message.

• Telegram's privacy mode is on by default, so groups need it off and the bot re-added, or the bot made admin.
• The gateway refuses anyone not on an allowlist or paired, which you should keep.
• Exec approval asks for a yes or no before risky commands, whatever model is running.

Since v0.23.0, one gateway owns every session, so a model switch in Telegram follows the conversation onto your desktop too.

Prefer to learn it yourself? The AI Profit Boardroom walks through model routing inside the Agent OS on video.

Also On Our Network

• Agent Operatorsthe six-step setup order, from BotFather token to locked-down group chats
• AI Income Deskwhether a small business should run it, with salon, shop and crew examples
• AI Tool Verdictan 8.1/10 scorecard across setup, access, voice, groups and files
• agentos.guidefive copy-paste prompts that set up groups, pairing, briefs and voice from inside the Agent OS

Set a fast default, give one channel a flagship, and if you want my team to check the plan, book a free strategy session before you pick the model behind your Hermes Agent Telegram bot.