Hermes Agent For Accounting + Which AI Models To Run: Real GoldieBench Scores For Local And Cloud Bookkeeping Jobs

Hermes Agent for accounting lets you choose the model for every bookkeeping job, and the first choice to make is not which model scores highest but whether your client paperwork should leave your machine at all.

Below are the GoldieBench scores I read on the live leaderboard and local-models pages on 8 November 2026, matched to the places Hermes lets you pin a model.

Each pairing names a real small-business job and what the model does in it, and none of this is tax or financial advice.

Want my team to choose models for your own workflows? Book a free SEO strategy session.

What GoldieBench Scores, And What Bookkeeping Needs Instead

The GoldieBench methodology sends every model the same fixed prompt, once, and the output must be a single self-contained HTML file.

Julian scores each result from 0 to 10, and the leaderboard ranks models by their average.

The task list is games, simulations, visuals and pages, not invoices or statements.

So the scores tell you how strong a model is at building something in one shot.

They don't tell you how accurately it copies a figure off a scanned receipt, and nothing on the site claims they do.

Read them as a guide to overall strength, then test the model on ten of your own documents.

Where Hermes Lets You Pin A Model For Finance Work

• The profile's main model: each Hermes profile has its own config, so a separate "books" profile can run a different model from your marketing agent.
• Every cron job: the cron docs resolve a job's model as per-job pin first, then cron.model, then the main model at the moment the job fires.
• A fleet default: set cron.model once, and switching your chat model never touches the month-end jobs.
• A locked choice: hermes cron create --pin freezes the current main model onto a job.

That last point matters for accounting: a month-end pack built by one model in October and another in November is harder to compare.

The Cloud Leaderboard, Read 8 November 2026

These are the top general models on the GoldieBench models page, with the averages exactly as shown.

Fusion is listed under OpenRouter, and the Hermes providers page covers OpenRouter among its cloud APIs.

Kimi K3's card says "1M context, tuned for long-horizon agent work", and GLM-5.2's says "1M context, open weights".

Long context is useful when one prompt has to hold a month of intake rows and a payment export together.

Note how close the middle of the table is: three of these models sit within 0.17 of each other, so the bench alone won't split them for a summary job.

The Local Leaderboard, For Books That Stay On Your Machine

The local models page only lists models that run fully on your own hardware, with nothing leaving the machine.

The Hermes providers page reaches local Ollama through its Custom Endpoint flow and lists privacy-sensitive work and offline use as what it's best for.

Agents-A1 and Gemma-4 12B Coder run through Ollama on that page, while Qwable 5 27B Coder runs through Apple MLX.

The gap is large: Qwable leads the local board at 7.14, and the next model sits at 4.83.

The 32k Context Warning For Small Local Models

The Hermes providers page has a warning every local set-up should read.

A small local model can choke on Hermes itself: all the default tools plus the system prompt may not fit in a 32k window, and llama.cpp-style servers then send back an empty stream.

Two ways out, per that page: hand the job fewer tools (the docs show -t file,web) or give the model more context when you load it.

For a receipts job that only needs to read files and write a workbook, a trimmed toolset is no loss.

Pairing: A Sole Trader's Receipts, Kept Local

1. A self-employed plumber runs a "books" profile pointed at a local model through Ollama.
2. Receipts arrive by email and land in a folder; the agent reads each PDF and adds a row to the workbook.
3. Anything it can't read clearly is marked UNSURE rather than guessed.
4. The plumber checks the rows each Friday, then hands the folder to the accountant.

Here the model choice is driven by privacy first, and the local board shows the trade-off in plain numbers.

A copy-and-flag job asks far less of a model than a one-shot game build, so a lower bench average can still be enough here.

Only the plumber's own Friday check can confirm that, which is why the check stays in the routine.

Pairing: A Bookkeeping Practice's Month-End Pack

1. A two-person practice keeps one Hermes profile per client.
2. Each profile's month-end cron job is pinned to one strong cloud model, such as Claude Opus 5 at 8.27 or GPT-5.6 Sol at 8.16.
3. The job builds a pack of unmatched items and missing documents, with every figure citing its source file.
4. Chase-up emails are saved as drafts, because Hermes approvals cover risky commands, not outgoing mail.
5. The bookkeeper reviews each pack before anything is posted.

A practice that wants client files on its own hardware would use the local route instead and accept the lower ceiling.

What The Payment Servers Add, Whatever The Model

Stripe, Square and PayPal all appear in the Nous-approved MCP catalog as vendor-hosted servers.

The MCP config reference lets you mark them trust: untrusted, so every write call asks first, and exclude refund and delete tools.

That guardrail sits in Hermes, not in the model, so it holds whichever model you pin.

There's no catalog entry for Xero or QuickBooks, so ledger work stays with a person.

Three Checks Before You Pin A Model To A Finance Job

1. Run it on ten real documents from last month and count every wrong figure.
2. Check it marks unclear fields instead of filling them in.
3. Check it keeps to "draft only" when a job touches a client.

A model that passes all three on your paperwork beats a higher average on someone else's test.

Keep the ten documents as a fixed test set, and run it again whenever you change the pinned model.

If a new model scores better on GoldieBench but worse on your set, your set wins.

What No Benchmark Can Tell You About Your Books

GoldieBench shows which models build well in one shot, and the local page shows what you give up by staying offline.

Neither can tell you whether a model reads your suppliers' layouts correctly, so your own test set is the final word.

If you want to see a hosted alternative scored the same way, read our Grok Bot for accounting model guide.

The AI Profit Boardroom trainings show model testing on real jobs, and the members there share what held up.

Also On Our Network

• Agent Operatorsthe eight-step books profile build, from receipts mailbox to month-end cron
• AI Income Deskwhether a small firm should use it, with five owner use cases and the risks
• AI Tool Verdicta nine-criteria scorecard for bookkeeping work, with strengths and limits
• agentos.guidefive copy-paste bookkeeping prompts and a weekly rhythm for your Agent OS
• aiprofitboardroom.comthe complete Hermes Agent for Accounting guide on the AI Profit Boardroom blog

Pick one finance job, test two models on your own documents, and book a free strategy session or join the AI Profit Boardroom if you want help choosing models for Hermes Agent for accounting.