LM Studio

The video is in Italian; this page is the written version in English.

If Ollama is the command-line engine, LM Studio is the dashboard: a desktop application to search, download, try and compare models in a few clicks. For anyone who dislikes the terminal it is the natural way in.

What it does

It searches and downloads models straight from Hugging Face without leaving the app, with a built-in chat to try them immediately. Before downloading it shows how much space a model takes and whether your machine can handle it — valuable information that saves a lot of wasted attempts.

It also exposes a local server compatible with the OpenAI API, exactly like Ollama, so anything you build on top is interchangeable.

The extra card on Apple Silicon: MLX

Here is the real difference. Besides GGUF, LM Studio serves models in MLX, Apple’s framework optimised for the unified memory of M-series chips. On the same machine with the same model, MLX is generally faster.

There is another detail that matters for agent builders: tool calling — the model’s ability to invoke external tools in a structured way — works reliably. That is why it gets picked as the backend for local agents such as Goose, where the reliability of tool calls is everything.

Ollama or LM Studio?

Not an exclusive choice: they coexist happily on the same machine, and many people use both.

	Ollama	LM Studio
Interface	command line	desktop app
Formats	GGUF	GGUF and MLX
Model discovery	know the name and pull it	visual search with filters
Automation	excellent, all scriptable	possible, but built for interaction
Tool-using agents	model dependent	more reliable

In practice: LM Studio to explore, work out what runs well on your machine, and serve agents; Ollama for scripting and services that must start on their own.

A practical tip

Use LM Studio for the selection phase — download three models, try them on the same prompts, watch speed and quality side by side. That is exactly the personal evaluation the benchmarks lesson argued for, and here it takes no code at all.

Using the local server from other programs

Beyond the interface, LM Studio exposes a server compatible with the OpenAI API: the part that makes it useful to developers too.

base address  →   http://localhost:1234/v1
key           →   any string (it is not checked)
model         →   the one loaded in the application

Three notes that save time: the model must already be loaded, because the server does not load on demand the way Ollama does; the name is the one shown in the interface, not the source repository’s; and if the calling program runs in a browser you need explicit permission for cross-origin requests, or the call is blocked before it starts.

When it is not the right choice

Worth saying so nobody wastes time: it is a desktop application, and on a machine without a screen it is the wrong tool.

Situation	Better
Headless server, automatic startup	Ollama as a service
Interactive use on your own laptop	LM Studio
Automation, scripts, scheduled jobs	Ollama
Local agents with tool calling	LM Studio, more reliable

Not an exclusive choice: they coexist on the same machine, on different ports. The criterion is simple — interactive and visual on one side, automatic and unattended on the other.

What to look at when picking a model in the interface

The built-in search shows dozens of variants of the same model and the temptation is to take the first. Three numbers decide for you.

File size, the real constraint: it must fit in memory with headroom, not at the limit. LM Studio shows an estimate and flags models that will not fit — the most useful piece of information in the whole interface.

Quantization, which you find in the name: Q4_K_M for daily use, Q5 or Q6 if you have memory to spare. Between a large heavily compressed model and a small lightly compressed one at equal gigabytes, the first usually wins.

Format: on Apple Silicon, when an MLX version exists it is generally faster than the GGUF one for the same model. On other platforms the question does not arise.

A time-wasting mistake: downloading three variants of the same model to compare them. Better to download one and try it on your real prompts — the difference between Q4 and Q5 shows far less than the difference between two different families.

Comparing two models on the same prompt

This is where the interface earns its place: the personal evaluation the benchmarks lesson argues for, with no code at all.

The method, a quarter of an hour:

1. Have twenty real prompts ready, the ones you would actually use, hard ones included
2. Load the first model and run them, pasting the answers into a file
3. Load the second — same prompts, same parameters, otherwise you compare two things
4. Reread the answers side by side counting real errors, not impressions

What almost always emerges: the differences are not where you expected. Two models a few points apart on benchmarks can behave in opposite ways on your task, and the one that looks better on paper may be the one that ignores the format you asked for.

Keep that file: when a new model appears, ten minutes tell you whether switching is worth it.

In short

Concept	In one line
What it is	Desktop app to download, try and serve models
MLX	Apple Silicon optimised format: faster on M chips
Tool calling	Reliable: the recommended backend for local agents
API	Local OpenAI-compatible server, like Ollama
Best use	Exploring and comparing models without touching a terminal
Local server	OpenAI compatible, but the model must be loaded first
When not	On a headless machine you want Ollama, not a desktop app
Picking a model	File size, quantization, format: in that order
Comparison	Same prompts and same parameters, answers side by side