Hermes Agent WhatsApp + Which AI Models To Run Behind It: Voice Notes, Photos And Reports Matched To Real GoldieBench Scores

Hermes Agent WhatsApp does not care which AI model you run, so the model question is really about what your WhatsApp chats will ask the agent to do.

The bridge and the Cloud API only carry messages, voice notes and files between WhatsApp and the Hermes gateway.

The thinking happens in whichever model Hermes is set to, and that is where the scores below come in.

Every number on this page was read on the live GoldieBench models page and the local models page on 11 November 2026.

To plan your model stack with my team, book a free SEO strategy session.

Three Jobs In A WhatsApp Chat, Three Different Model Needs

• Voice notes: both routes transcribe them with a speech-to-text provider such as local faster-whisper, per the bridge docs. Your chat model only sees the text.
• Photos and receipts: the Cloud API docs say vision models read images directly, while non-vision models get an auto-generated text description.
• Replies and actions: answering, filing, drafting and running tools all come down to the main model.

So the first filter is simple: if staff will send photos, pick a model that handles images natively.

What GoldieBench Measures, And What It Does Not

The methodology page sets the rules: one fixed prompt per task, one shot, no iteration and no best of several tries.

Each result must be a single self-contained HTML file, and every score is Julian's own 0 to 10.

That is a build test.

No task on the bench measures how well a model answers a short WhatsApp message or reads a blurry receipt.

Use the averages as a broad capability signal, then test your own messages.

The Cloud Leaderboard, Read On 11 November 2026

Scores change as tasks are added, so check the page before you lock anything in.

Reading The Table For A WhatsApp Agent

The top seven sit within half a point of each other.

For quick chat replies on a phone, that gap matters less than speed and whether the model reads images.

Fusion and Hermes MoA are panels rather than single models, per their taglines, which suits a heavy weekly job better than a two-line reply in a group chat.

The 1M-context entries, MiniMax M3, Kimi K3 and GLM-5.2, earn a look when the agent reads long forwarded PDFs or a big job log before answering.

Keeping The Model On Your Own Machine

The bridge already keeps the WhatsApp connection on your computer, with no Meta account and no public URL.

If you also want inference to stay local, the local leaderboard is the page to read.

The page itself notes a small local model will rarely beat a frontier cloud model on an ambitious one-shot build.

For a plumber logging jobs by voice note, the bar is lower than a one-shot game build, so a local model may be enough, but try it on your own messages first.

Example: A Salon Group Chat On A Fast Model

1. The salon's staff group runs on the bridge, with require-mention switched on.
2. Stylists @mention the agent for a reorder list from a shelf photo.
3. That needs a vision-capable model that answers quickly, not a slow panel.
4. The owner checks the draft list and places the order herself.

A user can switch models mid-chat with /model, and the gateway docs say that choice survives gateway restarts for that session.

Example: An Accountant's Weekly Summary On A Top Scorer

1. A cron job reads the week's forwarded receipts and drafts a summary.
2. The cron docs let one job pin its own model and provider with hermes cron edit <job_id> --provider <provider> --model <model>.
3. She pins this one job to a model near the top of the table and leaves daily chat on a faster one.
4. The summary is delivered to her WhatsApp as a draft for review.

On the Cloud API, that delivery fails with error 131047 if she has not messaged the bot in the past 24 hours, because Hermes has no template support yet.

Thinking Level Matters As Much As The Model

The cron docs also let one scheduled job pin its own reasoning effort, from none up to ultra, separate from the model pin.

That means a heavy weekly report can think at high while a routine morning reminder runs at minimal.

Levels a model does not support are clamped by the provider, so a higher pin never breaks the job.

For WhatsApp, where most replies are short, a low thinking level on chat plus a higher one on the big scheduled job is a sensible split.

A no-agent cron job, which only runs a script and delivers its output, makes no model call at all.

A Simple Model Plan By Route

For the support side of the same agent, see our Hermes customer support model guide.

What No Model Choice Fixes

A better model does not remove the bridge's ban risk; the docs advise a dedicated number and no bulk messages.

It does not add group support to the Cloud API route, which is direct messages only in v1.

And it does not change the allowlist: the gateway still refuses anyone not listed or paired.

Pick the route first, then the model.

Want to compare setups with people running them? Join the AI Profit Boardroom.

Also On Our Network

• Agent Operatorsthe exact commands and .env lines for both WhatsApp routes, plus the five errors to check first
• AI Income Deskwhich WhatsApp route a plumber, salon, accountant or shop should pick, with the risks
• AI Tool Verdicttwo scorecards rating the bridge and the Cloud API out of 10
• agentos.guidefive copy-paste prompts that make WhatsApp the front door of your Agent OS

Choose the route, then match a model to each job, and book a free SEO strategy session or join the AI Profit Boardroom if you want help picking models for Hermes Agent WhatsApp.