An LLM degree is a Master of Laws, from the Latin Legum Magister. It is a postgraduate law qualification, usually one year long, taken by people who already hold a first law degree and want to specialise or to practise in another country's legal system.
That answers the question as most people mean it. There is a second answer, and it is probably why you landed on a page published by a technical imprint. In machine learning, LLM stands for large language model, the thing sitting behind ChatGPT and every other chat box that has appeared in your software over the last two years.
Three letters, two fields that have never met. What follows covers the law degree quickly and the model properly, because the second meaning is what most people are chasing when they type the phrase.
What does LLM degree stand for in law
LL.M. stands for Legum Magister. The doubled L marks a Latin plural, which is why the degree carries two of them and why almost nobody punctuates it correctly on the first attempt.
The qualification sits above a first law degree. In the United States that means after the Juris Doctor. Elsewhere it follows an undergraduate law degree. Most programmes run a year full time and specialise in a single area such as tax or international arbitration among others.
Jurisdiction is the most common reason to take one. A lawyer qualified in one country uses an LL.M. to learn another country's system, which is exactly where the degree gets misread. The University of Texas School of Law states it plainly: an LL.M. does not make foreign lawyers eligible to practise law in the United States. Each state sets its own admission rules, and Texas and New York among others allow LL.M. holders to sit the bar under specific conditions.
This collision also explains the search results around this page. A query about LLM degree rankings returns law school league tables. A query about LLM training returns articles about GPU clusters. Two industries that have never once been in the same room, sharing an acronym, with one search engine quietly trying to work out which of you is asking.
If the law degree was what you came for, that is the honest summary and the rest of this page will be of no use to you. Everything below concerns the other one.
The other LLM: what a large language model is
A large language model is a program that predicts the next piece of text, trained on so much writing that the prediction turns out to be useful for nearly anything you can express in words.
That definition sounds far too small for what these systems do, and it is accurate anyway. Everything the model appears to know lives in its parameters, meaning the billions of numbers adjusted during training. Ask it something and it emits one token at a time, where a token is roughly three quarters of a word, each token conditioned on every token before it.
Hugging Face's LLM course describes them as a subset of natural language processing models characterised by their size and their training data, able to handle a wide range of language tasks with minimal task-specific training. That final clause is the part that reordered the industry. The older approach trained a separate model for each job. A large language model takes the job description at runtime and has a reasonable attempt at it.
The architecture underneath almost all of them arrived in 2017, when Vaswani and colleagues published Attention Is All You Need. The paper introduced the Transformer, an architecture built on attention alone, dispensing with recurrence and convolutions entirely. Every model you have used descends from that paper.
For readers who want the machinery below the API surface, LLM Systems Engineering works through tokenization and the transformer architecture in detail.
How an LLM is made, step by step
LLM training runs in stages. They differ enormously in cost, and confusing them is behind most of the wrong assumptions people bring to the topic.
| Stage | What happens |
|---|---|
| Data collection | Text is gathered at web scale, then filtered and deduplicated |
| Tokenization | A vocabulary is built, turning text into the integer IDs the model actually consumes |
| Pretraining | The model learns next-token prediction across the whole corpus. This is where the compute budget goes |
| Post-training | Supervised fine-tuning on curated examples, then preference tuning to shape how it answers |
| Evaluation | Benchmarks and human review decide whether the result is shippable |
Pretraining is the stage that gets written about, and it is the one you are least likely to perform. A frontier run occupies thousands of accelerators for weeks. The capability the model holds at the end of it is broad and largely shapeless.
Post-training is where the assistant you recognise gets built. Supervised fine-tuning teaches the model the shape of a helpful answer, and preference tuning nudges it toward responses that human raters preferred. The special tokens marking whose turn it is in a conversation acquire their meaning here, a detail that stays invisible until it breaks, covered in how special chat tokens are trained.
The step-by-step framing hides one thing worth stating. At no point does a person write down a rule that the model then follows. Behaviour is a statistical consequence of the data and the objective, which is why a model can be excellent at your task on Monday and confidently wrong about it on Tuesday for reasons nobody in the building can reconstruct.
Is ChatGPT an LLM?
Yes, with a clarification worth carrying around. ChatGPT is a product built on top of a large language model. The GPT model serving your account is the LLM. ChatGPT is the interface and the safety layer around it.
The distinction becomes load-bearing the moment you build something. Calling the API gives you the model with none of the product attached. No memory between calls, and no web search unless you wire one up yourself. Teams find this out at the demo, when a prompt that behaved beautifully in the chat window comes back through the API as a complete stranger.
The same holds for Claude and Gemini among others. Each name covers a family of models plus a set of products built over them, and the marketing rarely bothers to separate the two. If you are starting from zero, Claude AI for Beginners walks through the difference between a model and a product before it asks you to write anything.
There is no LLM degree for large language models
If you came looking for a university qualification in large language models, the honest answer is that no accreditation body has minted one. Machine learning masters programmes carry LLM material inside them, and the field reliably outruns any syllabus on a two-year revision cycle.
What exists instead is a stack of skills that hiring managers actually test for:
- Prompting and context engineering: the highest leverage per hour invested, and the least respected
- Evaluation: building a harness that tells you whether a change helped, which almost nobody does properly
- Retrieval: getting the right documents in front of the model at the right moment
- Fine-tuning: adapting a model once prompting has genuinely run out of road
- Inference economics: latency and cost, where most production projects quietly die
Notice what is absent from that list. Training a foundation model from scratch appears nowhere, because vanishingly few organisations need to. When to build an LLM covers where that line sits. For the prompting layer, AI Prompt Engineering treats prompts as a maintained system with a lifecycle.
None of this requires a degree. It requires an evaluation set you wrote yourself and a tolerance for being wrong in public.
A local LLM teaches more than a syllabus
The fastest route into these systems is to run one on your own machine. A local LLM, meaning a model whose weights sit on your hardware and whose inference never leaves it, makes the abstractions concrete in an afternoon.
Open-weight models in the 7 to 14 billion parameter range run on a recent laptop with enough memory. You will watch tokens appear one at a time, feel the context window fill, and see quality fall away when you quantise too aggressively. Reading about any of that conveys perhaps a tenth of what an hour of it does.
A private LLM carries a second argument beyond learning. Nothing you send it is transmitted anywhere, which settles a class of questions that otherwise need a legal review. For regulated data, that property is frequently the entire reason a project is permitted to exist. Confirm the property before you rely on it: some local tooling ships with telemetry or cloud fallback enabled by default, and the guarantee holds only once you have verified the model is answering from disk.
Be realistic about the ceiling. A model you run locally will trail a frontier API model on hard reasoning, and that gap is real. It narrows for well-defined tasks, and for learning purposes it does not matter at all.
So the summary is short. An LLM degree is a Master of Laws and always has been. A large language model is a next-token predictor that turned out to be startlingly general. If you searched the phrase meaning the second one, close the university prospectus and download a set of weights this evening. Whichever LLM you were after, one of them will admit you tonight without an application fee.