All Things AI · Season 2 · The Models · Episode 3
GLM is an open model you run on your own machine, and it is the one to reach for when the job is sustained writing rather than extraction. Free, offline, and strong enough at long-form drafting that most agents stop paying for that particular task.
What it is
A family of open models, published rather than sold, downloaded and run on hardware you own. Same arrangement as any local model: no account, no usage charge, no request leaving the building.
Where it earns a separate episode is that it is comparatively strong at producing long, coherent, structured text. That is a genuinely different capability from following a tight instruction, and it is the thing most open models are weakest at.
Why long-form is a different skill for a model
It is not simply more of the same, which is why a model can be excellent at one and mediocre at the other.
Holding a shape across a whole piece
Producing eight hundred coherent words means keeping track of what has already been said, not repeating it, and building toward something. Many models produce excellent paragraphs that do not add up to a piece, and the failure only appears at length.
Transitions
The connective tissue between sections is where writing either flows or reads as assembled. It is also the first thing to go when a model is working section by section, which is why single-pass drafting matters.
Not losing the register
A consistent voice across a long piece is harder than a good voice in a paragraph. Drift is the common failure, where the opening sounds like one person and the close sounds like a different one.
What to use it for
Anything where the output is a piece rather than an answer.
- Full first drafts from your own outline and material, produced in one pass.
- Long explanatory content, which is the highest-value thing an agent can publish and the most tedious to write.
- Rewriting something at a different length or for a different audience without losing what it said.
- A second opinion on anything another model drafted, because two open models cost nothing to run and disagree usefully.
Where it is weaker
It is less rigid about format than a model tuned for instruction-following, so for extraction work where you need identical structure across fifty items, Hermes is the better tool. Running both is free and there is no reason to pick one permanently.
As with every local model, it has no current information and no way to obtain any, and it will not always signal that clearly.
And it is not the most capable model available for genuinely hard reasoning. Long-form fluency and difficult analysis are separate things, and it is strong on the first.
The pairing that makes both better
The reason to run two open models rather than one is that it costs nothing and the second opinion is frequently better than either.
Give the same request to both. Where they agree, you can be more confident. Where they diverge, the divergence itself is informative, because it usually marks the place where your request was ambiguous rather than where one model was wrong.
This is a technique that is only practical because local models are free. On paid services nobody runs everything twice, which means nobody gets the benefit.
How you run it
Identical to any local model, and if you already have one running this is fifteen minutes.
- The same free application that manages local models will list it with a download button.
- Pick a size that fits your memory. Larger models write better and the trade is speed, so this is the one place where sizing up is worth considering if the machine allows it.
- Give it your outline rather than asking it to invent one, because the structure is where your thinking lives.
- Ask for the whole piece in one pass instead of section by section.
What to try in the first week
Two exercises, and both are designed to show you the thing it does that a hosted chat window does not.
Give it a real outline of yours, five or six bullets, along with your actual notes, and ask for the entire piece at once. Not an introduction, not a section, the whole thing. Then read it as a piece rather than as paragraphs and notice whether it holds together, because that is the specific test this model is meant to pass.
Then run the identical request through a second local model and read both. The comparison teaches you where your request was vague, which is almost always the real problem when output disappoints. It costs nothing but the time to read twice, and it is the fastest way to get better at asking.
Where this goes wrong
Three failures.
Asking it to invent the structure. It will produce a reasonable one and the piece will read as competent and impersonal. Five minutes of outline from you is what makes the output yours.
Building in fragments. Section-by-section assembly produces repetition and missing transitions, which is exactly the weakness this model exists to avoid. One pass.
Using it for extraction. It is less rigid about format than the instruction-tuned alternative, and if you need identical structure across many items you are using the wrong one of two free tools.
Why publishing long explanatory content is worth the trouble
Worth saying plainly, because this model is most useful for a kind of content most agents have stopped making.
Short content competes on volume and disappears in a day. Long explanatory content is the opposite: it takes real effort, it is the thing people actually search for, and a good piece keeps working for years. The reason almost nobody produces it is not that its value is unclear. It is that an eight-hundred-word explanation of something takes an evening, and evenings are scarce.
It also happens to be exactly what the newest search layer rewards. A page that answers one question completely, in clear quotable sentences, is what a model draws from when it composes an answer. Short posts have nothing for it to use.
So the capability and the opportunity line up unusually well here. The thing that was too expensive to produce is the thing that is now most worth having, and the tool that makes it affordable is free.
How it sits next to the others
Against hosted services it trades some polish and some reasoning capability for being free, private, permanent, and unable to change underneath you.
Against Hermes it is the drafting half of a pair rather than a competitor. Hermes reads and extracts, GLM writes at length, and the sensible setup for most agents is both installed and used for what each is good at.
Against a model trained on your own writing, it is the base rather than the alternative. An adapter carrying your cadence has to sit on top of something, and a strong long-form model is the right thing for it to sit on.
What it costs
Nothing, permanently, with no subscription or usage charge. Disk space and an evening, and if you already run one local model, considerably less than an evening.
Questions agents actually type
What is GLM used for?
Sustained long-form writing rather than extraction. It holds a shape across a whole piece, produces transitions, and keeps a consistent register at length, which is the specific thing most open models are weakest at.
Should I run more than one local model?
Yes, because they are free. Giving the same request to two and reading both tells you where your request was ambiguous, which is usually the real problem when output disappoints. Nobody does this on paid services because nobody runs everything twice.
Is GLM or Hermes better for a real estate agent?
They are a pair rather than a choice. Hermes reads documents and extracts with rigid formatting. GLM drafts at length. Both are free, so the sensible setup is both installed and each used for what it is good at.
Can a free local model really write well enough to publish?
For long explanatory content, yes, provided you supply the outline and your own material and then correct the result. What it cannot do is invent your argument or supply specifics only you know, and those are what make writing worth reading.
Season 1 is Claude Code for Real Estate Agents, filmed and on its own hub. All three seasons: AI for real estate agents.