Hermes Agent For Lawyers + Which AI Models To Run, With Real GoldieBench Local And Cloud Scores

Hermes agent for lawyers puts the model choice in the firm's hands, and for client files the right first filter is where the model runs, not where it ranks.

Below are the scores I read on the live GoldieBench models and local-models pages on 9 November 2026, matched to legal jobs and to the places Hermes lets you set a model.

None of this is legal advice, and a qualified lawyer reviews every output whichever model produced it.

Want help choosing models for your firm's workflows? Book a free SEO strategy session with my team.

What These Scores Measure, And What They Don't

The GoldieBench methodology says every demo is one prompt, one shot, with a single HTML file as the output.

Julian scores each result from 0 to 10, and the ranking is the average across every task a model was scored on.

The tasks are games, simulations and pages.

Nothing on the site tests contract review, legal research or citation accuracy, and I won't pretend it does.

Use the bench as a signal of general strength, then test any model on a handful of your own dummy matters before trusting it.

The First Filter: Local Or Cloud

Hermes Local Models runs open models on your own machine, and its docs say nothing leaves the computer once the model is downloaded.

The Hermes FAQ adds that requests go only to the provider you configure, so a cloud model means your prompts reach that company.

For privileged material, that makes the local board the starting point and the cloud board a partnership decision.

The two boards answer different questions, which is exactly how the local-models page frames itself.

The Local Board, Read 9 November 2026

The local models page lists models it describes as strictly on-device, with the averages below as shown.

Qwable 5 27B Coder leads by a wide margin, and its runner is MLX, which points to Apple hardware.

The page also lists Laguna XS 2.1 at 3.93, but its own row says it was benched through Poolside's API with local support still to come, so I've left it out of a client-data shortlist.

Ornith 1.0 shows as pending, so there's no score to quote.

What A Local Model Should Do In A Small Firm

Even the top local score sits well below the cloud leaders, so give local models jobs that are easy to check.

• Intake summaries: a solo practitioner's questionnaire becomes a summary and a missing-documents list, saved as a file.
• Text extraction: the PDF skill pulls text from the PDF's text layer, and the model only has to list dates and clauses.
• Checklists from skills: a property practice's completion procedure, saved as a skill, is applied to each new matter.
• Reminder drafts: a morning due-soon list built from a deadlines file the lawyer keeps, with every date verified by her.

Each of those has a source document next to it, so a lawyer can check the output line by line.

Avoid open-ended jobs on a small local model, such as asking it to judge the strength of a claim.

That is a lawyer's call in any case, and weaker models are more likely to sound confident while being wrong.

Keep prompts narrow, name the file, and ask for the clause number beside every point.

The Cloud Board, Read 9 November 2026

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

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

Hermes MoA's card describes a panel of frontier models merged by a chair, so it sends work to several cloud providers at once.

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

Long context matters when one prompt has to hold a long agreement plus its schedules.

When A Cloud Model Is Worth Considering

Some firm work involves no client data at all.

Rewriting a public-facing practice-area page, summarizing a published statute, or drafting an internal training note are examples.

For those, a stronger cloud model may be the sensible choice, set up through the Hermes providers you already use.

Keep the rule simple: no client names, facts or documents go to a cloud model unless the firm has approved that provider.

Firms that do approve one can look at community masking plugins in the plugin catalog, but check each entry carefully; one of them says it is fail-open on errors.

Where Hermes Lets You Set The Model

• Per profile: a separate legal profile can run a local model while your other agents use something else.
• Per cron job: the cron docs resolve a job's model as its own pin first, then cron.model, then the main model when the job fires.
• A locked choice: hermes cron create --pin freezes the current main model onto a job, and the docs say the agent cannot point a job at a different model.
• Self-hosted endpoints: the providers page covers Ollama and vLLM for a firm server.

Pin the deadline-reminder job to your local model so a chat-model switch never sends that file to a cloud provider.

A Pairing Plan For A Two-Lawyer Firm

1. Install Hermes on the office machine and download Qwable 5 27B Coder if the hardware is Apple silicon, or one of the Ollama models on the local board otherwise.
2. Create a legal profile that uses only that local model.
3. Pin every cron job in that profile to the same model.
4. Run intake summaries and PDF extractions on dummy matters and have a lawyer mark every error.
5. Only if the firm approves a cloud provider, set up a second profile for public, non-client drafting.

I wrote a similar split for bookkeeping in Hermes Agent for accounting, and the same logic carries over.

What The Bench Can't Tell A Lawyer

• Whether a model reads a 200-page bundle accurately; that depends on the text layer and on reading in chunks.
• Whether a model cites the right clause; check every citation yourself.
• Whether a provider meets your professional duties; the Hermes docs make no SOC 2, HIPAA or ISO claims.
• Whether a date is right; cron reminders repeat what you entered, so verify each one.

For the Agent OS and the walkthroughs, the AI Profit Boardroom has the full setup.

Also On Our Network

• Agent Operatorsthe nine settings I'd change before Hermes reads a single client file
• AI Income Deskfive small-firm jobs it can prepare, and the risks an owner keeps
• AI Tool Verdictan 11-row scorecard that rates privacy high and legal integrations low
• agentos.guidefive copy-paste prompts for a private legal desk in your Agent OS
• aiprofitboardroom.comthe complete Hermes Agent for Lawyers guide on the AI Profit Boardroom blog

Start on the local board, keep a lawyer on every output, and join the AI Profit Boardroom to test models for your own Hermes agent for lawyers.