LM Studio vs Ollama: we ran a local AI model on a normal laptop
We ran the same free 9B coding model in LM Studio and Ollama on a Windows laptop without a gaming GPU. Private and free, but slow, and less than half of our coding checks passed.
Can a normal laptop run its own AI model, with no internet and no bill? We installed the two best-known free apps for this, LM Studio and Ollama, on a Windows laptop without a gaming graphics card. Then we gave the same small AI model our secret coding test. Both apps work, and your questions stay on your laptop. But the model was slow and passed less than half of our checks, while a cheap online model passed all of them.
What are LM Studio and Ollama?
An AI model is a big file full of numbers. Apps like ChatGPT run that file on their own computers. Ollama and LM Studio let you run a model file on your own computer instead. This is often called a "local LLM": a large language model that runs locally. Your questions and the answers then never leave your laptop.
- Ollama is a small program that runs in the background. It has a simple chat window and a command line. You pick a model by name, like "ornith:9b", and it downloads it for you. It is open source: anyone can read and reuse its code, under the MIT licence.
- LM Studio is a full desktop app. You can search models, see how big each one is, chat, and change settings with buttons. It is free to use, but its code is not public.
Both can also act as a local "server". That means other programs on your computer, like a coding tool, can talk to the model as if it were an online AI. We used that server for our test.
Our laptop and our test
We used a normal Windows 11 laptop from 2026. It has an Intel Core Ultra X7 358H chip with built-in Intel Arc B390 graphics, and 32 GB of memory, of which 31.4 GB is usable. There is no separate graphics card. Most gaming PCs have one, and AI models like it best.
The model was Ornith 1.0 9B, a free coding model that both apps offer. "9B" means it has 9 billion numbers inside. We used the same "4-bit" version in both apps: the numbers are stored in a shorter form, so the file is only 5.6 GB. In Ollama that is "ornith:9b". In LM Studio we used the official file from Hugging Face, a big website for AI models. They are the same model and size, but not the exact same file.
Then we gave it our secret coding test, the same one we use for online models. It has three jobs: fix the money maths in a small program, write a bill splitter, and write a CSV reader. Each job came in one message, and the model had to answer with the full code. We ran every job three times per setup. Hidden tests checked the code: 186 checks in total. Each answer could take at most 15 minutes. You can read more in how we test AI models.
We tested three setups: Ollama as it comes, Ollama with the graphics chip switched on, and LM Studio as it comes. In all three, we gave each answer a working space of 16,384 tokens, which is more than Ollama gives by default.
Installing both
Both apps installed on Windows without an account.
- Ollama 0.40.6 took about 2.8 GB on disk. It also put itself in the Windows Startup folder, so it starts every time you turn on your computer. Downloading the model was one command:
ollama pull ornith:9b. - LM Studio 0.4.26 took about 0.8 GB for the app and 1.3 GB for its engines. Our first download of the model, from LM Studio's command line, stopped at 7%. A bit later, the app finished the download by itself.
The model itself was 5.6 GB in both. So count on about 8 GB of free disk space for one app and one small model.
The GPU surprise
The biggest difference showed up before we asked a single question.
LM Studio looked at our laptop and found the built-in graphics chip, also called the GPU. It reported 17.88 GiB of shared memory for it. Then it picked its "Vulkan" engine by itself. Vulkan is a standard way for programs to use a graphics chip.
Ollama found the same chip, but did not use it. Its own log said: "dropping integrated GPU; to enable, set OLLAMA_IGPU_ENABLE=1". So on its default setting, Ollama ran the model on the main chip, the CPU, only. Ollama's GPU page does not mention this setting.
To switch it on, we started Ollama by hand from the command line, with that setting. Then Ollama put the whole model on the graphics chip, and it got about 60% faster. There is no button for this in the app; you need to set it yourself. Afterwards we set Ollama back to normal.
How fast is it?
Speed for AI models is measured in tokens per second. A token is a short piece of a word.
- Ollama, as it comes (CPU only): about 11.6 tokens per second.
- Ollama, graphics chip on: about 18.6 tokens per second.
- LM Studio, as it comes: about 17.7 tokens per second.
That is faster than most people read, so it sounds fine. With the graphics chip in use, both apps were about equally fast. That fits, because both build on llama.cpp, a free program that runs these model files.
But for coding, the model first "thinks" in hidden text, and that thinking can be thousands of tokens long. Each answer to our money job took 2.6 to 3.4 minutes in LM Studio. On a typical job, Claude Haiku 5.5 answered in about 15 seconds online.
How good was the code?
Out of 186 checks, the model passed:
- 61 on Ollama as it comes (33%),
- 85 on Ollama with the graphics chip on (46%),
- 66 on LM Studio (35%).
These gaps look big, but we would not read too much into them. A few things seem to explain them:
- Luck. With only three tries per job, results jumped a lot. In LM Studio, the money job scored 26, then 15, and then 0. That last code got stuck in an endless loop: it kept running and never finished.
- Time. On Ollama's slow setting, 5 of 9 answers ran past our 15-minute limit. Those count as zero.
- Settings, maybe. Ollama's version of the model comes with a short instruction: "Be concise, correct, and direct". It also sets the "temperature" to 0.6; a lower temperature makes the answers less random. In LM Studio, we kept the app's own settings. There, the model wrote more: about 13.6 thousand tokens for the bill splitter, against about 8.4 thousand in Ollama with the graphics chip. Several LM Studio answers only just made it under 15 minutes. We did not test the settings one by one, so this may only be part of the story.
For comparison, from earlier tests on the same set: Claude Haiku 5.5 passed all 186 checks for about 2 cents in total, and GPT-6 Luna passed 184. The bigger open model Qwen 3.8 27B passed only 62 of 135 when we tested it online. You could run that one at home too, on a strong computer. So it is not just our laptop: the open models we tried also had a hard time with these jobs.
Why one job never finished
The CSV job turns a text file with commas into a table. It scored zero in every setup on our laptop. Mostly that was not because the code was wrong, but because there was no code.
The model kept thinking. In some tries it hit our 15-minute limit. In others it used up all 16,384 tokens of working space with thinking, and had no room left for the answer. One try in Ollama thought for 868 seconds and wrote 15,746 tokens, and the answer was empty. One try in LM Studio did write code first, but then kept thinking inside its answer until the space was full. That code did not even load.
You could give the model more room and more time. But at about 18 tokens per second, an answer this long already takes around 15 minutes on our laptop. Online models finished this job in seconds to a few minutes.
How much memory you need
The model needed about 6 GB of memory. Ollama reported 5.9 GB, and Windows showed 6.6 GB for LM Studio's engine. The built-in graphics chip has no memory of its own; it borrows from the same 32 GB. With a browser, a chat app and a few other programs open, at one moment only about 1 GB was still free.
So here is a rough rule from our test. For a small 9B model, you want 16 GB of memory at the very least. If you want to keep other apps open too, 32 GB is better. Bigger models need much more: Ollama's library lists the 27B model above as an 18 GB download. That would not have fit next to our other apps.
Privacy, licences and cost
Privacy. With a local model, your chats stay on your computer in both apps. LM Studio's privacy page says your messages, chats and documents are never sent from your computer. It does check for updates, and that sends your app version, system and internet address. Ollama's privacy policy, from March 2026, says it does not collect your prompts. It does collect some usage data without the content, like the app version and how many requests you make. Note that Ollama also offers cloud models, with ":cloud" in the name. Those run on other computers, so your text does leave your laptop.
Licences. Ollama is open source under the MIT licence, which lets anyone use and change it. LM Studio is free, but under its own terms. You may use it at home and inside your company. You may not use it to sell a service to others.
Cost. Both apps and the model are free. You pay for power and for the laptop. There is no bill per question, unlike online models. But online can be very cheap too: our whole test cost about 2 cents with Claude Haiku 5.5. See our AI API price comparison.
Which one should you pick?
- New to this, or a laptop without a graphics card? Pick LM Studio. It found and used our graphics chip by itself, and you can do everything with buttons.
- A developer who wants a small background service? Ollama is lighter and open source. On a laptop like ours, set it to use the built-in graphics, or it will be slower.
- You need code that works? On a normal laptop, a small local model is not there yet. Use an online model, or a computer with a strong graphics card and a bigger model.
Local AI is great for private chats, for trying out models, and for working on a plane. Just don't expect it to beat a good online model on a laptop like ours.
Questions people ask
Does Ollama run AI models locally?
Yes. In our test, Ollama downloaded the 5.6 GB model Ornith 1.0 9B and ran it on our laptop, with no internet needed for the answers. But Ollama also offers cloud models, with ":cloud" in the name. Those run on other computers, so your text leaves your laptop.
What are the disadvantages of Ollama?
On our laptop, Ollama did not use the built-in graphics chip on its default setting. Its log said so, and we had to start it with the setting OLLAMA_IGPU_ENABLE=1 to make it about 60% faster. There is no button for this. It also adds itself to Windows Startup. And the small model we used passed only part of our coding test.
Is LM Studio totally free?
LM Studio is free to use, at home and inside your company, under its own terms. It is not open source, and its terms do not let you use it to sell a service to others. The models you download are free too, each under its own licence.
Is LM Studio really private?
With a local model, LM Studio's privacy page says your messages, chats and documents are never sent from your computer. It does check for updates, which sends your app version, system and internet address, and model searches are sent without your name. We did not check its network traffic ourselves.
Is local LLM better than Chat GPT?
For privacy and working offline, yes: nothing leaves your laptop. For quality and speed on a normal laptop, no. Our local model passed 85 of 186 coding checks at best, while Claude Haiku 5.5 online passed all 186, much faster. We did not test ChatGPT itself here.
How much RAM for local LLM?
A small 9B model at 4-bit needed about 6 GB of memory on our laptop. The built-in graphics chip borrows from that same memory. We would say 16 GB is the least, and 32 GB is better if you keep a browser and other apps open. Bigger models need much more.
What does local LLM mean?
LLM stands for large language model, the kind of AI behind chat apps. Local means it runs on your own computer instead of on a company server. Apps like Ollama and LM Studio download the model file and run it for you, so your questions stay on your computer.
Is it worth running local LLM?
It is worth it if you want privacy, want to try models, or work offline. It is not worth it yet for real coding work on a laptop without a strong graphics card. In our test the model was slow and often ran out of time. A cheap online model did the whole test for about 2 cents.
How we checked
On 20 and 21 October 2026 we installed Ollama 0.40.6 and LM Studio 0.4.26 on the site owner's laptop. It runs Windows 11 on an Intel Core Ultra X7 358H with Intel Arc B390 graphics and 32 GB of memory. Other everyday apps stayed open, as on a normal working day. We used Ornith 1.0 9B at 4-bit (Q4_K_M) in both apps. In Ollama we used "ornith:9b" from its library. In LM Studio we used the official file from Hugging Face, and we checked its checksum, a code that shows the file is complete and unchanged.
We sent our secret coding test v2 to each app's local server: three jobs, three tries each, 186 checks. Each job was one request, with a working space of 16,384 tokens and a limit of 15 minutes. Hidden tests graded the code, and code that hung for more than 2 minutes scored zero. Speeds and token counts come from each app's own numbers.
Ollama ran two ways: as it comes, and with OLLAMA_IGPU_ENABLE=1 to use the graphics chip. Both Ollama setups used the settings that come with the model in Ollama. LM Studio used its own defaults; we did not change them.
Things went wrong a few times, and we redid that work. The laptop went into standby twice, which made it slower. Once, a helper script of ours ran both apps at the same time. Once, our script lost its connection to LM Studio. We threw away every answer that was hit and ran it again.
We also read Ollama's privacy policy and GPU page, and LM Studio's privacy page and terms. The online scores come from our earlier runs of the same test. The charts are our own, and none of the images was made by AI. This was one laptop and one model; other laptops, models and settings will give other results.