All Things AI · Season 2 · The Models · Episode 2
Hermes is an open model you download and run on your own computer. It is tuned to follow instructions closely rather than to chat, which is exactly what you want from something processing documents. It is free, it runs offline, and the file never leaves your laptop.
What it is
A family of open models, meaning the model itself is published and you can download it and run it on hardware you own. There is no account, no per-use charge, and no request travelling anywhere.
The specific thing that distinguishes Hermes from other open models is what it was tuned for. A base model is trained to continue text. Hermes has been further trained specifically to follow instructions accurately and to stay inside the format and constraints it was given, which is a different objective from being a good conversationalist.
Why instruction-following is the property that matters
This sounds like a technicality and it is the whole reason to choose one open model over another for this kind of work.
It stays in the format you asked for
Ask for a list of five items with a specific structure and you get a list of five items with that structure, rather than a friendly paragraph explaining what it found followed by four items in a different shape. When you are processing fifty documents, format consistency is the difference between usable output and an afternoon of cleanup.
It respects constraints
Told not to speculate, it is comparatively good at not speculating and saying it could not find something instead. No model does this perfectly. The difference between one that mostly does and one that mostly does not is large in practice.
It does not try to be interesting
A model tuned for conversation improvises, adds context, and offers observations. For document work all of that is noise, and a model that does the narrow job you asked for is considerably more useful than a charming one.
What to use it for
The profile is document work, repetitive work, and anything you would rather not send anywhere.
- Reading long documents and pulling out specific things, with a citation requirement.
- Extracting structured information from unstructured text, which is the single most useful thing it does.
- Comparing documents against each other and reporting differences.
- Any repetitive processing where the same operation runs across many items and you want identical treatment each time.
- Anything at all involving a client’s material, because this is where running locally stops being a preference and becomes the obvious answer.
Where it is weaker
It is not as capable as the largest hosted models on genuinely difficult reasoning, and pretending otherwise sets people up for disappointment. For a hard multi-step problem, a paid service is still better.
It knows only what it was trained on and has no connection to anything current. It cannot look something up, and it will not always make that limitation obvious.
It is less fluent as a writer than models tuned for that. It will produce accurate, well-structured, slightly flat prose. For extraction that is ideal. For something with your name on it, draft elsewhere or with an adapter trained on your own writing.
How you run it
Straightforward, and the whole thing is an evening at most.
- Install a free application that manages and runs local models. Several exist and they present a list with a download button next to each.
- Choose a size that fits your machine. Memory is the constraint, not processor speed, and starting smaller than you think is the right call.
- Confirm it is fast and pleasant before doing anything real with it. A model that is too large for the hardware will work and be miserable, and that experience is what makes people give up.
- Then turn off your internet connection and use it, once, deliberately, because seeing that work is what makes the privacy argument concrete.
Where this goes wrong
Three failures, and the first accounts for most abandoned attempts.
Downloading the largest model available. It will load, it will be painfully slow, and the conclusion drawn will be that local models are not viable. They are. That one was too big for that machine.
Expecting current information. It has none and no way to get any. Anything that depends on what is true today comes from your files or from a service that is genuinely online.
Assuming private means correct. Running locally answers where the data went. It says nothing about whether the output is right, and a confident wrong answer produced privately is exactly as wrong.
How it sits next to the others
Against the hosted services it trades capability for privacy, permanence, and cost. It is less able on the hardest tasks and it cannot be changed, repriced, or discontinued by anyone.
Against other open models it is the one to reach for when the job is instruction-shaped. Where the job is long-form drafting, GLM is the better companion, and running both is free so there is little reason to choose permanently.
The natural arrangement for most agents is Hermes as the default for anything document-shaped or client-attached, with a paid service kept for the small number of tasks that genuinely need more.
What to try in the first week
Three exercises, and the first one is the one that convinces people.
Disconnect from the internet completely and use it for twenty minutes on real work. Everyone reads that a local model runs offline and nobody quite believes it until they watch it happen with the connection off. It reframes what this is in a way no explanation does.
Take a long document and ask it to extract every date, deadline, and time period, with the page for each. This is the task it is unusually good at, it is genuinely tedious by hand, and the output is immediately checkable, so you learn how much to trust it within minutes rather than weeks.
Then ask it something it cannot know, deliberately, and see what it does. Something current, or something about a specific place, or something that requires looking anything up. Watching how it behaves when it does not know is more useful than any amount of watching it succeed, because that behaviour is what you are relying on when you are not checking.
Why an open model matters beyond privacy
The privacy argument is the obvious one and it is not the most important one over a long period.
A model on your disk is a fixed thing. It behaves identically in three years, it cannot be updated underneath a workflow you built, and it does not stop existing because a company changed direction. Anyone who has built a process on a hosted service has had the experience of it behaving differently one morning for no reason they were told about.
For anything you intend to rely on, that stability is worth a meaningful amount of raw capability, and most agents undervalue it until the first time it bites them.
What it costs
Nothing, permanently. No subscription, no usage charge, no account. The only cost is disk space, an evening of setup, and the eventual memory decision when you next replace the laptop.
Questions agents actually type
What is Hermes AI and how is it different from ChatGPT?
Hermes is an open model you download and run on your own computer rather than a hosted service you send requests to. It is tuned specifically to follow instructions and stay in the format you asked for, which suits document work better than conversation does.
Is Hermes free?
Yes, permanently. The model is published, you download it, and it runs on hardware you already own with no subscription, no per-use charge, and no account. The only real cost is disk space and an evening of setup.
What is Hermes best used for by a real estate agent?
Reading long documents and extracting specific things, comparing documents against each other, and any repetitive processing where you want identical treatment across many items. Anything involving a client’s material belongs here rather than on a hosted service.
Why does my local model give slow or poor answers?
Almost always because the model chosen is too large for the machine’s memory. Start with a smaller one, confirm it is fast and pleasant, and only move up if you actually need to. Memory is the constraint rather than processor speed.
Season 1 is Claude Code for Real Estate Agents, filmed and on its own hub. All three seasons: AI for real estate agents.