CodeHero A local AI model is an open model that runs on your own hardware, so the work it reads never leaves the building. Ollama, vLLM and LM Studio are three common ways to run one. The limit is memory: the model has to fit, and larger models need much more of it.
Why would a business run a model locally?
To keep the material the model reads inside, to work without depending on an outside service, and to pay for hardware once instead of for every request.
When a cloud AI service does a step of the work, it receives the context that step needs. For some businesses that is fine; for others — client files, health data, internal know-how — keeping the model inside is the simpler answer.
What is the difference between Ollama, vLLM and LM Studio?
Ollama runs open models on your own computer with a simple command and an API; vLLM is a library for serving models with high throughput, made for a server with graphics cards; LM Studio is a desktop app with a local server.
Ollama is a quick way to start: you pull a model and use it from apps, coding agents or your own code, through its API, which also works with OpenAI- and Anthropic-style clients.
vLLM is built for serving: it manages memory carefully and batches incoming requests, so many requests can share the hardware, and it offers an OpenAI-compatible API server. It suits a dedicated server that several people or agents use at once.
LM Studio runs on macOS, Windows and Linux, uses GGUF models (and MLX models on Apple Silicon), and serves them through a local server with an OpenAI-compatible API — locally or across your network.
| Ollama | vLLM | LM Studio | |
|---|---|---|---|
| What it is | A way to run open models on your computer | A library for serving models with high throughput | A desktop app for running local models |
| Suits | A first start, one computer | A dedicated server that many people or agents use at once | A desktop or laptop, also shared on the network |
| How apps connect | Its API, also for OpenAI- and Anthropic-style clients | An OpenAI-compatible API server | A local server with an OpenAI-compatible API |
What hardware do you need?
For the local models we have tested and recommend for this work, at least 96 GB of memory for the model itself.
Memory decides what fits: on a computer with unified memory it is the system memory; on a desktop with graphics cards it is the cards’ memory added together. The hardware page lists examples, from computers with large unified memory to desktops with several graphics cards.
Are local models as good as the big cloud models?
Not at everything. The largest cloud models are stronger at long, complex work; good local models do well on clear, well-described tasks.
That suggests a mixed setup: a large model plans the work and checks it, while a local model takes the clear, well-specified pieces. The better the plan and the checks, the more of the work a smaller model can do well.
Where does CodeHero fit?
CodeHero connects to local models through Ollama, vLLM, LM Studio and other OpenAI-compatible servers, next to the cloud AI companies, and you choose per role and per ticket which model does what.
The factory itself runs on your own computer or server; with a local model, the work that model reads stays there too. Independent checks, repairs and the record work the same whichever model does the step.