I mentioned last week that I’m spending a lot of time with LLMs recently. I’m poking and prodding them to test their ‘understanding’ (inverted commas necessary there!) of phonology, in particular with non-standard speech and dialects.

And you’d be forgiven for thinking I’m just tapping my prompts into ChatGPT, Claude, Gemini or the other big commercial concerns. Mention AI, and those are the names people come up with. They’re the all-bells-and-whistles web-facing services that get all the public fanfare and newspaper column inches.

The thing is, that’s not all there is to Large Language Models. There’s a whole world of open source (or the slightly less open ‘open weights’) models out there. Some of them offshoots of those big names, while others less well-known. But you can download all of them to run offline on any reasonably-specced laptop.

LMStudio – LLMs on your laptop

Meet LMStudio – the multi-platform desktop app that allows you to install and interrogate LLMs locally. It all sounds terribly technical, but at its most basic use – a custom chatbot – you don’t need any special tech skills. Browsing, installing and chatting with models is all done via the tab-based interface. You can do much more with it – the option to run it as a local server is super useful for development and testing – but you don’t have to touch any of that.

Many of the models downloadable within LMStudio are small models – just a few gigabytes, rather than the behemoths behind GPT-5 and other headline-grabbing releases. They feature the same architecture as those big-hitters, though. And in many cases, they are trained to approach, even match, their performance on specific tasks like problem-solving or programming. You’ll even find reasoning models, that produce a ‘stepwise-thinking’ output, similar to platforms like Gemini.

A few recent models for download include:

• Qwen3 4B Thinking – a really compact model (just over 2gb) which supports reasoning by default
• OpenAI’s gpt-oss-20b – the AI giant’s open weights offering, released this August
• Gemma 3 – Google’s multimodal model optimised for use on everyday devices
• Mistral Small 3.2 – the French AI company’s open model, with vision capabilities

So why would you bother, when you can just fire up ChatGPT / Google / Claude in a few browser clicks?

LLMs locally – but why?

Well, from an academic standpoint, you have complete control over these models if you’re exploring their use cases in a particular field, like linguistics or language learning. You can set parameters like temperature, for instance – the degree of ‘creativity wobble’ the LLM has (0 being a very rigid none, and 1 being, well, basically insane). And if you can set parameters, you can report these in your findings, which allows others to replicate your experiments and build on your knowledge.

Small models also run on smaller hardware – so you can develop solutions that people don’t need a huge data centre for. If you do hit upon a use case or process that supports researchers, then it’s super easy for colleagues to access the technology, whatever their recourse to funding support.

Secondly, there’s the environmental impact. If the resource greed of colossal data centres is something that worries you (and there’s every indication that it should be a conversation we’re all having ), then running LLMs locally allows you to take advantage of them without heating up a server farm somewhere deep inside the US. The only thing running hot will be your laptop fan (it does growl a bit with the larger models – I take that as a sign to give it a rest for a bit!).

And talk of those US server farms leads on to the next point: data privacy. OpenAI recently caused waves with their suggestion that user conversations are not the confidential chats many assume them to be. If you’re not happy with your prompts and queries passing out of your control and into the data banks of a foreign state, then local LLMs offer not a little peace of mind too.

Give it a go!

The best thing? LMStudio is completely free. So download it, give it a spin, and see whether these much smaller-footprint models can give you what you need without entering the ecosystem of the online giants.

Lastly, don’t have a laptop? Well, you can also run LLMs locally on phones and tablets too. Free app PocketPal (on iOS and Android) runs like a cut-down version of LMStudio. Great for tinkering on the go!