You can train a model on your own writing so the draft already sounds like you. You can run one on your own laptop so a file never leaves it. You can build the tool you need in an afternoon instead of renting one that almost works. That is what this is.
Most agents have only seen the smallest version of this
You typed something into a box. It typed a paragraph back. The paragraph was fine and generic, it did not sound like you, and you closed the tab and went back to work.
That is the version almost every agent has met. It is also the least interesting thing any of these models can do, and it is why so much of the industry has decided AI is a novelty that writes mediocre listing copy.
The gap between that box and what is actually sitting on the table right now is enormous. Not in a few years. Right now, on a normal laptop, with no developer and no budget that would make you flinch.
This hub exists to close that gap. Three seasons, twenty-one episodes, each one explained end to end: what it is, what it takes to run, what comes out, and what it grows.
It can read the files you already have
You do not have to paste anything. A model can be pointed at the folder on your computer and work from what is in it. The export, the report, the notes you took last Tuesday, the twelve documents attached to a deal. It reads them where they sit.
That single change turns a chat window into something that knows your business, because your business is already sitting in those folders.
It can remember
A chat forgets when you close it. That is a product decision, not a limit of the technology. Rules, preferences, corrections, and past work can persist and be searched across everything you have ever told it.
The practical version: you stop re-explaining your own business every morning. You correct something once and it stays corrected. Two computers can hold the same memory and reconcile to each other, so the knowledge is identical whichever one you sit down at.
It can be trained on your writing
Feed a model enough of your own work and it learns your cadence. Not a style prompt. An actual adapter trained on what you have already written, so the draft comes back in your rhythm and you change one line instead of rewriting the page.
This is the difference between a tool that produces content and a tool that produces your content. The LoRA episode in Season 2 walks through exactly how it is done and what it takes.
It can run entirely on your own machine
Local models run offline on a normal laptop. No subscription, no service to go down, and nothing sent anywhere. For anything that touches a client document, this is the answer to a question most agents have never thought to ask about where their data goes.
It is also free, permanently, once it is set up.
It can be more than one
One chat window is one worker. You can run a group. One that plans, one that writes, one that checks facts against a source, one that keeps the brand consistent, each with its own instructions, all working the same job and handing off to each other.
This is the single biggest jump in what is possible, and it is the one almost nobody has shown agents.
What this actually grows
Every capability on this site is here because it moves one of four things, and it is worth being blunt about which.
It gives you back hours. The admin that surrounds a transaction, the follow-up you meant to send, the document you meant to read, the video you meant to film. That is the most obvious gain and the least interesting one.
It makes you consistent. Most agents are excellent in bursts. A system that holds your rules and your voice makes the tenth week look like the first week, which is the whole game in a business built on repeat and referral.
It makes you findable. Search is splitting into layers, and the newest one is being the answer inside an AI rather than a link on a page. Almost no agent has done anything about it.
And it changes what you are competing with. An agent who runs this themselves is not doing the same job faster. They are doing a job that used to require staff.
What agents actually search, and why this hub is built the way it is
This site already ranks for brokerage and business questions, and those pages keep that work. This hub is built for the questions it was not answering: how agents use AI, which AI to use, whether it can run on your own computer, whether it can sound like you, and what it can be made to do beyond writing a paragraph.
Answer engines want a short true answer first and the detail after. So every page here leads with the answer in the first forty words, then teaches the thing properly underneath. No page competes with another page for the same question, which is the mistake that keeps most agent sites invisible.
There are three layers to being found now, and they are not the same discipline.
Search
The results page. Still the largest source of intent, still worth doing properly, and still where most agents stop.
Answers
The box at the top that answers without a click. Winning it is a structural problem, not a keyword problem: a direct answer, in the first lines, on a page that owns that question and nothing else.
Generative results
Being the answer when someone asks an AI directly who to work with, or what to do next. This is the newest layer, the least contested, and the one with the shortest window.
Season 1 — Claude Code
Filmed, live, and on its own hub. Claude Code is the deepest single track on this site: skills, automations, a ten-part build series, and the workflows that run a business day to day. If you want one thing to go all the way through, start there.
Season 2 — The Models
Eleven episodes, one per model. What it is, what it costs, what it is genuinely better at than the others, and what you would reach for it to do. You do not need all of them. You need to know which one to pick up, and that is a five-minute answer per model instead of a year of trial and error.
Season 3 — The New Models
Ten episodes, one model each, and every one of them new to this hub since Season 2 went up. Not a rehash of the eleven you already met. What each model’s parts do, what they cost, and what you would reach for them to handle.
What each model is actually for
Eleven models, eleven short answers. This is the section to read if you have been opening whichever one you heard of first and wondering why the results are underwhelming. Full episode behind each one.
Claude Code
The one that works inside your actual files and runs multi-step jobs rather than answering questions. It reads a folder, writes to it, runs the steps, and reports back. This is the deepest track on the site and it has its own hub, its own build series, and its own season.
Reach for it when the job has steps. Not write me a paragraph, but read these twelve files, pull what matters, produce the document, and save it where it belongs. That is a different category of task from anything a chat window does.
Grok
Fast, current, and comfortable with a lot of context at once. Useful when you want a draft quickly from material you already have rather than a careful long workshop. Cheap enough to run in batches, which matters more than it sounds when you are producing at volume.
Reach for it when you need volume and speed and you already have the raw material. Twenty drafts from twenty sets of notes is an afternoon on other tools and a coffee break on this one.
Hermes
An open model you can run on your own machine. It follows instructions closely, which is what you want from something working on documents rather than chatting. Nothing leaves the laptop, and there is no monthly cost once it is set up.
Reach for it when the document is a client document. The question is not whether a service is trustworthy. The question is whether the file needs to leave your desk at all, and here it does not.
GLM
Another open model, strong at long-form drafting, and a good second opinion when one model keeps producing the same shape of answer. Running two open models and comparing costs nothing but time.
Reach for it when the first draft came back flat. A second open model with a different training gives you a genuinely different attempt rather than the same answer reworded, and both are free to run.
DeepSeek
An open model published under the MIT license, with weights anyone can download, and priced through its own service at a fraction of what the other paid models charge. It is far too large to run on a laptop, so the practical version is hosted: we run it through NVIDIA’s API on free trial credits, which keeps the file on NVIDIA’s servers rather than DeepSeek’s and the cost at nothing.
Reach for it when the cost per draft is the thing deciding whether you run the job at all. A hundred follow-ups, a batch of recaps, a season of captions: the price stops being a reason to do fewer of them.
ChatGPT
The one that is genuinely good on a phone, with voice, while you are between appointments. Not the one to reach for when the work involves your own files, but the most convenient by a wide margin when your hands are full.
Reach for it between appointments. Voice, phone, hands full, thirty seconds to think out loud and get something usable back before you walk into the next thing.
Gemini
Sits where a lot of agents already live, inside documents, mail, and calendar. The value is proximity: the material is already there, so nothing has to be moved for it to be useful.
Reach for it when the material is already in your mail or your documents and moving it anywhere else is the actual obstacle. Proximity beats capability more often than anyone admits.
LoRA
Not a chatbot. A small adapter trained on your own writing that makes another model produce your cadence. This is the piece that turns generic output into your output, and it is trained on work you have already done.
Reach for it once, and then it improves everything else you run. This is infrastructure rather than a tool, which is why it sits in Season 2 as a model even though you never talk to it.
Memory
Also not a chatbot. The layer that lets any of the above hold your rules, corrections, and past work across months, and search all of it on every request instead of starting from nothing each time.
Reach for it the second time you find yourself explaining the same preference to a machine. That moment is the signal, and most agents ignore it for a year.
Local models generally
The category rather than a product: any model running on hardware you own. Free after setup, offline, private, and increasingly good enough that the gap to a paid service does not matter for most day-to-day work.
Reach for them when the answer to what does this cost per month needs to be nothing, or when the answer to where does this go needs to be nowhere.
Cursor
Built to see a whole project of files at once rather than a single document. Agents use it for the same reason developers do: it can hold everything at once and change things across all of it consistently.
Reach for it when the work spans many files at once and they need to stay consistent with each other. That is a common shape of problem and almost nothing else handles it well.
The ten new models
Season 3 in one place. Ten models, each with several parts, each with a full episode, and every one of them new to this hub since Season 2 went up.
Meta Muse
A family, not one product: Muse Spark (the model behind it), Muse Code (a terminal coding agent), Muse Glimmer (a free, open-weight version you can run yourself), the Muse app (a personal agent that checks with you before it sends, buys, or publishes anything), and Muse for Small Business (works your numbers and drafts campaigns, same approval rule).
Qwen
Qwen3.8 covers text, images, and video, with a free open-weight version you can download and run on your own machine. Qwen-Image-2.1 both creates and edits images in one model, including backgrounds you can see through. A Qwen 4 preview is already in training.
Kimi K3
The model itself is built for long documents and long-horizon research with a million tokens of context. Its “Agent Swarm” mode runs hundreds of subagents on one job at once, and its “OK Computer” mode builds a working website or a slide deck from a plain-English brief.
MiniMax
H3 (also called Hailuo) generates video with its own sound already built in. Music 3.0 writes, arranges, and performs a complete song in one pass. The M-series text models hold a long context and can operate a desktop computer directly.
NVIDIA Nemotron
An open family with a fast small model (Lightning) for high-volume work, larger reasoning models (Super and Ultra) for harder jobs, a model that reads video, audio, and images together (Nano Omni), and dedicated helpers for transcription and telling speakers apart.
Xiaomi MiMo
A flagship reasoning model (Pro) and a faster everyday version (Flash), both reading text, images, video, and audio together, plus a text-to-speech line that can clone a voice from a short sample or design a brand-new one from a written description.
Google Veo and Omni
Veo 3.1 generates video with real sound, not silent clips with music added after. Gemini Omni lets you edit a video in plain words while it keeps the same people and places consistent. Both live inside Google Flow and Google Vids, where most agents already have an account.
LTX
LTX-2.5 is open-weight, free to run on your own hardware, and can string several camera angles into one generation while keeping the same subject across the cuts. The earlier LTX-Video model remains a fast, free way to test a single clip before committing to more.
Mistral OCR
A document-reading service built for PDFs, forms, tables, and signatures across 170 languages, returning the text along with exactly where it sat on the page. Vibe, Mistral’s own chat and work assistant, sits alongside it for the rest of the job.
OmniVoice
A free, run-it-yourself voice model that clones or designs a voice from a short sample across hundreds of languages, plus a desktop app built around it for dictation and dubbing without an account or a monthly bill.
How the three seasons fit together
Season 1 is depth on one thing. Claude Code, filmed, with a full build series behind it. If you want to go all the way through a single stack, that is the track.
Season 2 is orientation. Eleven models, eleven short answers, so you stop guessing which one to open. Most agents pick whichever they heard of first and then conclude the whole category is disappointing.
Season 3 is what came after Season 2. Ten models, one per episode, every one of them new to this hub since Season 2 went up, each broken into the parts that actually matter: what it does, what it costs, and what you would reach for it to handle.
Where to start this week
Pick the job you are most tired of. Not the most impressive one. The one that eats a Tuesday evening every week without ever being interesting.
Then find the episode that covers it and run just that. One capability, working, beats five half-configured tools every time.
- Tired of writing? Start with LoRA, training a model on your own words, Season 2, Episode 7.
- Tired of re-explaining yourself to a chat window? Start with memory, Season 2, Episode 8.
- Worried about where client documents go? Start with local models, running one on your own computer, Season 2, Episode 9.
- Drowning in long documents? Point one of the models above at the file and ask it what matters.
- Not sure which model to open at all? Season 2, Episode 10.
What this costs
Worth answering plainly, because the assumption that this is expensive is the reason a lot of agents never start.
The open models are free. Hermes, GLM, and anything else you run on your own machine cost nothing per month, cost nothing per use, and keep working whether or not you keep paying anyone. The only cost is the afternoon it takes to set up the first one.
The paid services sit in the range of a single monthly subscription each, which is less than most agents already spend on tools they use once a week. You do not need several. Season 2 exists so you can pick one or two deliberately instead of collecting them.
Training a model on your own writing is a one-time job on hardware you already own. Video, voice, and translation are metered, so those are the ones to watch, and the episodes say so where it applies.
There is nothing to buy from us at any point. Every episode is built so you can run the thing on your own accounts and keep it.
This teaches you to run it, it does not sell you a stack
There is no software here to buy from us. Every episode is built so you can go and do the thing yourself, on your own accounts, with your own files, and keep it when you are done.
That is deliberate. An agent who rents a system is one price change away from losing it. An agent who can run the system owns something.
Questions agents actually type
What can AI actually do for a real estate agent?
Draft in your own voice from your own files, remember your rules across months, run on your own computer without sending anything out, read long documents in seconds, produce video at volume, work inside software you already pay for, and make you findable inside AI answers. You decide and you send. It does the work around that.
Do I need to know how to code?
No. These models are instructed in plain English, which is the entire reason this became available to agents rather than to engineering teams. Season 3 is built around that assumption.
Which AI should a real estate agent start with?
Whichever matches the job in front of you, which is a different answer for writing, for reading documents, for working with files on your computer, and for anything you want kept private. Season 2 gives a five-minute answer per model so the choice takes minutes instead of months.
Can AI run on my own computer?
Yes. Local models run offline on a normal laptop with nothing sent anywhere, no subscription, and no service that can go down. Season 2, Episode 9 covers what it takes and what it is good for.
Can a model actually sound like me?
Yes, if it is trained on your own writing rather than told to imitate a style. That is a specific process using work you have already produced, and the result is one pass you correct rather than a draft you rewrite.
Can AI use more than one tool at once?
Yes. A group of models can work a single job together, each with its own instructions and its own role, handing work to each other. It behaves less like a tool and more like staff.
What is GEO and why does it matter now?
Getting found inside AI answers instead of on a results page. It is a different discipline from traditional search, it is very early, and early is the whole advantage.
Will AI replace real estate agents?
It replaces the work around the job, not the judgment inside it. Negotiation, advocacy, and the moment a client is anxious are not tasks. What changes is how much of everything else you are still doing by hand.
How much does AI cost for a real estate agent?
Open models you run yourself are free permanently, with no per-use charge and no subscription. Paid services are roughly one monthly subscription each and you need one or two, not several. Video, voice, and translation are metered and worth watching. Training a model on your own writing is a one-time job on hardware you already own.
Everything you publish is yours, and Fair Housing applies to it exactly as it always has.