Where do you start if you want to run and train AI on your own computer?
Start by running a ready-made open model with a free tool such as Ollama or llama.cpp, because that shows you what your memory allows before you spend anything.
8 min read
Train and run your own AI
Run models on your own computer, choose hardware, and train, adapt and test your own models.
DGX Spark, Ryzen AI Max or Mac Studio: which small AI workstation makes sense?
All three hold 64 GB or more of shared memory in a small box, so they load models a single consumer graphics card cannot.
6 min read
How do you build your own language model, and which path fits your goal and budget?
Building your own model can mean anything from a free laptop exercise to a run costing hundreds of thousands of dollars, so choose by goal.
5 min read
What hardware should you buy to run AI at home in late 2026?
Buy for memory first: 16 GB runs small models, 32 GB runs today's best 27-billion-parameter coding models, and 128 GB reaches models above 100 billion parameters.
6 min read
How much do power, heat and noise really cost when you run AI at home?
Electricity is the small part: four hours a day costs roughly CA$1 to CA$3 a month on a shared-memory desktop in Québec, and up to about CA$11 on a large graphics card PC.
5 min read
Should you pretrain a model, fine-tune one or use RAG, once you count the real costs?
For almost every practical task, use RAG for facts that change and fine-tuning for consistent behavior; together they cost little.
4 min read
How do you train a small language model on a rented GPU, step by step?
Rent a machine by the minute, connect over SSH, clone Karpathy's nanochat, run a short test model for about $3, then run the full script for about $64 on eight H100 GPUs.
5 min read
What does it really cost to train a language model in 2026?
At rental prices read on November 2, 2026, a GPT-2-grade model costs about $64, a 124-million-parameter nanoGPT reproduction about $1,500, and a capable 3-billion-parameter model about $880,000 at list price.
5 min read
How do you fine-tune a small open model yourself?
Pick a small open-weight model of 3 to 8 billion parameters, prepare a few hundred to a few thousand examples of the exact input and output you want, and train a QLoRA adapter with a tool such as Unsloth or…
4 min read
Mac or PC with a graphics card: which is better for local AI?
A PC graphics card is much faster for any model that fits inside its memory, and nearly all AI software targets it first.
5 min read
How do you train a tiny language model from scratch to understand how it works?
Start with Andrej Karpathy's nanoGPT example, which trains a character-level model on Shakespeare in minutes on an ordinary laptop, then move to his nanochat project, which runs every stage of a small chatbot…
4 min read
Which open coding models can you run on your own computer?
As of October 2026, the two strongest coding models that fit on one home machine are Alibaba's Qwen3.8-27B and Meta's Muse Glimmer-30B, both under the permissive Apache 2.0 licence.
3 min read
How do you turn a model checkpoint into a working app?
Package the weights in a safe format with a config, a version and a fingerprint, test that it gives the same numbers as the training checkpoint, then wrap it in a small server with input limits and a health…
4 min read
How do you diagnose a failed training run before buying a bigger machine?
Run four cheap checks: the first loss against what guessing should give, whether the model can memorize 16 examples, how large the gradients are and where the time goes.
5 min read
How do you train an image style and speech recognition on your own machine?
For images, train a small diffusion model, then add a low-rank adapter and show it about 30 pictures of the new style; for speech, teach a small network commands and test it on voices it has never heard.
5 min read
How much memory does an AI model need, and will it fit on your machine?
Multiply the parameters in billions by the bits per weight, then divide by 8, and you get the file size in gigabytes.
5 min read
What AI can really run on your phone in 2026?
A recent phone can run models of about 2 to 4 billion parameters at a useful speed, enough for summaries, rewriting, dictation and short questions.
5 min read
How do you teach a model to choose tools and make decisions?
Train on examples of request and tool call, test on wordings and values it has never seen, then have ordinary code validate each call and hold risky ones for a person.
4 min read
How do you teach a model a writing style without changing the facts?
Train on pairs where the facts vary and only the style repeats, include the odd cases such as missing values, then check every draft with a plain program that compares it to the source.
4 min read
When should you adapt a small language model, and when should you not?
Adapt a model to teach behaviour, such as a fixed output format, and use a lookup for facts.
4 min read
How do you fine-tune an open-weight model with LoRA?
Collect a few hundred examples of the exact input and output you want, then train a small LoRA adapter on top of a frozen open-weight model with Hugging Face's TRL and PEFT libraries.
4 min read
How do you search your own documents with embeddings, and prove that it works?
Turn each passage and each question into an embedding with a small open model, return the closest passages, then test it with questions whose right answer you know.
4 min read
Why can a model fit on your machine and still be impossible to train?
Running a model mostly needs room for its weights, while training also needs a gradient for every weight, the optimizer's two extra copies and all the activations that must be kept for the backward pass.
5 min read
Mixture-of-experts or dense, what does each design demand from hardware?
A dense model uses all its parameters for every token, while a mixture-of-experts model uses only a small fraction, so it gets more capacity for the same computing per token.
5 min read
What happens when a tiny language model learns from random weights?
It starts out knowing nothing, and its first loss proves it by matching the math of pure guessing, then it learns letters, then common words, then something shaped like sentences.
5 min read
How do you go from a classifier to finding objects in images?
A classifier gives one label per image, while a detector turns the same kind of network into a grid in which every cell says whether an object is centred there, where its box is and what it is.
4 min read
How do you know your model actually improved?
Compare old and new on the same untouched test set, count which examples each one alone gets right, put a confidence interval on the difference and repeat with other seeds.
4 min read
How do you build a dataset that teaches the job, legally?
Start from the decision the model must make, then collect examples of real use that you have the right to use, remove personal details and duplicates, and record where everything came from.
5 min read
How do you turn a local model into a server your apps can use?
Start a server program such as Ollama, llama-server or vLLM, and it will answer on a local address using the same request format as OpenAI's API, so most apps only need a new base URL.
4 min read
What are parameters, loss and learning, and what do they look like when you watch them change?
Parameters are the numbers a model is allowed to change, loss is one number saying how wrong it is, and learning is the repeated nudge of each parameter in the direction that lowers the loss.
4 min read
How do you train your first model on any laptop, with no GPU?
Install scikit-learn, load its built-in handwritten digits, and fit a small neural network on the CPU; on one processor thread this took under half a second and reached 96.9% accuracy on images it never saw.
4 min read
How do you build your own tokenizer for a language model?
Train a byte pair encoding (BPE) tokenizer: start from the 256 possible byte values, repeatedly merge the most frequent adjacent pair into a new token, and stop at your chosen vocabulary size.
4 min read
How do you prepare a dataset to train a language model from scratch?
Choose text whose licence fits what you plan to do with the model, then clean it, remove duplicates and personal details, shuffle it, set part aside for validation and store it in shards.
4 min read
What do Q4_K_M, GGUF, MXFP4 and FP8 mean, and which one should you download?
GGUF is a file format, while Q4_K_M, MXFP4 and FP8 are ways of storing each number in fewer bits.
5 min read
How do you tell whether a language model you trained is any good?
Check it at four levels: validation loss, reported as bits per byte, shows whether it learned to predict unseen text; fixed sample prompts show what that means in practice; public benchmarks compare it with…
4 min read
What hardware do you need to run AI models at home in 2024?
The deciding factor is memory: the whole model has to fit in your graphics card's memory (VRAM) or, on an Apple silicon Mac, in its unified memory, with some room to spare.
4 min read
What is Ollama, and how does it let you run AI models on your own Mac?
Ollama is a new free, open-source tool, first released as an early preview on July 8, 2023, that downloads and runs open large language models on your own computer with one short command, such as "ollama run…
4 min read
How do you run a language model on your own laptop with llama.cpp?
You build llama.cpp from its GitHub page, convert Meta's LLaMA model files into the program's own format, shrink them to 4-bit numbers, and then run the model from the terminal.
4 min read
What is llama.cpp, and how can a large language model run on a laptop?
llama.cpp is a small open-source program, released on GitHub on March 10, 2023 by Georgi Gerganov, that runs Meta's LLaMA language models in plain C/C++ on an ordinary computer's processor.
4 min read