Where do you start if you want to run and train AI on your own computer?

Running an AI model on your own computer and training one are two different jobs. Running (also called inference) means loading a model someone else trained and asking it questions. Training means changing a model's numbers, its parameters, until it does a job better. Running needs room for the model. Training needs that room several times over. This page puts the guide's articles on both in a sensible order, and answers the question most people ask first: what does my machine allow?

Step 1: run a model before you buy anything

The quickest way to learn what your computer can do is to run a ready-made open model on it. Free tools do the hard part:

• What is Ollama? explains the one-command tool most people start with.
• What is llama.cpp? explains the program underneath Ollama and many other apps.
• What do Q4_K_M, GGUF, MXFP4 and FP8 mean? tells you which file to download. For most people, a Q4_K_M file is the safe default.
• How much memory does an AI model need? gives the rule of thumb: parameters in billions, times bits per weight, divided by 8, gives the file size in gigabytes. Then add room for the conversation and the operating system.

Worked example: a 27-billion-parameter model in the Q4_K_M format, which llama.cpp measures at about 4.9 bits per weight, is 27 × 4.9 ÷ 8 ≈ 16.5 GB of weights, before the conversation and the system take their share. That is why the strongest open coding models that fit at home are roughly 18 GB downloads and call for a graphics card with 24 GB or a computer with plenty of unified memory.

Once a model runs, How do you turn a local model into a server your apps can use? shows how other programs can talk to it, and What AI can really run on your phone in 2026? covers the smallest devices.

Step 2: understand why training needs so much more memory

Training keeps more in memory than running does: a gradient for every parameter, the optimizer's extra copies, and the intermediate results needed to work backwards through the network. Why can a model fit on your machine and still be impossible to train? measured it on a 135-million-parameter model: full training took 5.1 times the memory of running it, and a LoRA adapter (a small set of extra numbers trained on top of a frozen model) cut the peak by two-thirds. That ratio is the reason most people adapt models with LoRA or QLoRA rather than training every parameter.

So what can 24 GB do?

The table summarizes what the guide's articles report. Memory sizes refer to what the graphics processor can really use; on shared-memory computers, that is less than the amount installed. The training minimums come from the documentation of Unsloth, an open fine-tuning library, and grow with longer examples.

Sources for the rows: What hardware do you need to run AI models at home?, What hardware should you buy to run AI at home in late 2026? and How do you fine-tune a small open model yourself? Check your own case with the memory rule above before buying anything: the speed of the memory and the software you plan to use matter too, as Mac or PC with a graphics card explains.

Step 3: the training series, in order

These fourteen articles build on each other. Each one was run on a real machine and reports what was measured, including what failed. Most run on an ordinary laptop.

1. How do you train your first model on any laptop, with no GPU?
2. What are parameters, loss and learning?
3. How do you build a dataset that teaches the job, legally?
4. How do you know your model actually improved?
5. How do you go from a classifier to finding objects in images?
6. What happens when a tiny language model learns from random weights?
7. Why can a model fit on your machine and still be impossible to train?
8. How do you search your own documents with embeddings, and prove that it works?
9. When should you adapt a small language model, and when should you not?
10. How do you teach a model a writing style without changing the facts?
11. How do you teach a model to choose tools and make decisions?
12. How do you train an image style and speech recognition on your own machine?
13. How do you diagnose a failed training run before buying a bigger machine?
14. How do you turn a model checkpoint into a working app?

The series has one lesson that runs through every article: a score is only worth something if the test was honest. Article 4 shows a common data mistake reporting 84% accuracy on pure random numbers, where the honest answer was 51%.

Step 4: adapting open models and building your own

When the series is done, these articles cover adapting real open models and, for the curious, building one from scratch:

• Should you pretrain a model, fine-tune one or use RAG, once you count the real cost? For almost every practical task: RAG (giving the model your documents) for facts that change, fine-tuning for consistent behaviour.
• How do you fine-tune an open-weight model with LoRA? and How do you fine-tune a small open model yourself?
• How do you tell whether a language model you trained is any good?
• How do you build your own language model, and which path fits your goal and budget?, with preparing a dataset, building a tokenizer and training a tiny language model from scratch.
• How do you train a small language model on a rented GPU, step by step? and What does it really cost to train a language model in 2026?

Step 5: hardware, only when you know what you need

Buy for memory first, and only after you have run models on what you already own:

• What hardware should you buy to run AI at home in late 2026?
• Mac or PC with a graphics card: which is better for local AI?
• DGX Spark, Ryzen AI Max or Mac Studio: which small AI workstation makes sense?
• Mixture-of-experts or dense: what does each design demand from hardware?
• How much do power, heat and noise really cost when you run AI at home?

What can a machine with 24 GB of memory actually run and train?

It runs mid-size open models well, including the strongest coding models that fit at home, in their 4-bit versions of roughly 18 GB. It can adapt open models with QLoRA up to about 27 billion parameters, or with 16-bit LoRA up to about 8 billion, which is how most people teach a model a format, a style or a set of tools. It cannot fully train anything large: full training needs several times the memory of running, so serious pretraining means renting GPUs by the hour, and the cost article shows what that adds up to. Before deciding a bigger machine is the answer, run the four cheap checks in How do you diagnose a failed training run before buying a bigger machine?: in nine broken models tested there, more GPU would have fixed none of them.