Two paths for artificial intelligence
As AI has grown more powerful, a quiet but important debate has taken shape over how these systems should be shared. On one side are closed models, offered only as a service through a company's own website or interface. On the other are open models, which anyone can download, inspect, and run on their own computers. The choice between them shapes who controls this technology.
Both approaches have serious supporters and real trade-offs. Understanding the difference helps make sense of much of the news about AI — including debates over safety, competition, and national policy.
What open source actually means
In software, open source has long meant that the underlying code is public. Anyone can read it, change it, and build on it. Open-source AI extends this idea to models. Instead of only offering answers through a paid service, developers release the model's trained parameters — the numbers that make it work — for others to use freely.
The definitions get blurry in practice. Some models release everything, including the training data and methods. Others release only the finished parameters under conditions. Purists argue about where the line truly sits, but for most people the key point is simple: an open model can be downloaded and run without asking permission from the company that made it. Like the software behind many blockchains, open AI projects rely on communities that anyone can inspect and join.
Why open models matter
Open models change the balance of power in several practical ways.
- Control — organisations can run models on their own hardware, keeping sensitive data in-house.
- Cost — after download, running a model locally can be far cheaper than paying per request.
- Customisation — developers can adapt and fine-tune a model for a specific task or language.
- Transparency — researchers can study how a model behaves and where it fails.
- Resilience — a business is not tied to one provider's pricing, rules, or availability.
For a hospital, a bank, or a government that cannot send confidential data to an outside service, these benefits are not luxuries. They are the difference between using AI at all and leaving it on the shelf. Smaller open models can even run on your own laptop or phone, a form of edge computing.
The safety debate
Openness cuts both ways. The same freedom that lets a researcher study a model also lets a bad actor misuse it. Once a capable model is downloaded and shared, it cannot be recalled. Critics worry this makes it harder to prevent misuse, from generating scams to producing harmful content.
Open models spread capability widely. The question is whether the benefits of transparency outweigh the risks of losing control.
Supporters counter that transparency is itself a form of safety. When many independent experts can examine a model, flaws are found and fixed faster than in a closed system that only its owner can inspect. They also argue that concentrating powerful AI in a few private companies carries its own risks. There is no settled answer, and reasonable people disagree. It is one of the central questions in the wider ethics of artificial intelligence.
The surprising quality of open models
A few years ago, open models lagged far behind the best closed systems. That gap has narrowed dramatically. Well-designed open models now handle many everyday tasks — writing, summarising, coding, answering questions — at a level that satisfies most practical needs, even if the very largest closed models still lead on the hardest problems.
This progress matters because it lowers the barrier to building with AI. A small startup, a university lab, or an individual developer can now create useful products without a giant budget, simply by downloading a capable model and adapting it. That accessibility is fuelling a wave of experimentation.
This wave matters beyond the technology world. When the barrier to building with AI falls, innovation is no longer confined to a few wealthy companies. Researchers in universities, developers in smaller countries, and hobbyists working evenings can all contribute, adapt models to local languages, and tackle problems the big providers may never prioritise. History suggests that this kind of broad access tends to produce unexpected and valuable results, precisely because so many different people are free to try. Open models also create a healthy check on the closed ones: when a capable free alternative exists, commercial providers must keep improving and pricing fairly, because customers always have somewhere else to go. That quiet competitive pressure may prove one of the most important consequences of the open-source approach, benefiting even those who never touch an open model themselves.
Meet the leading open models
New versions arrive constantly — check the labs' own announcements for the latest. The families below have consistently led the open field:
- DeepSeek — whose recent V4 generation became famous for matching the best closed models on reasoning at a fraction of the usual training cost.
- Alibaba's Qwen — the Qwen 3 family is known for strong multilingual ability and coding skill, and is hugely popular with developers.
- Meta's Llama — the fourth-generation Llama family is among the most widely downloaded open models in the world — a strong all-rounder.
- Google's Gemma — Google's lightweight open family, with Gemma 3 releases designed to run on modest hardware, including laptops and phones.
How do open models compare on benchmarks?
Independent benchmark comparisons in 2026 have told a striking story: averaged across many tests, the best open models now trail the best closed models by only a couple of percentage points — one widely cited comparison put the gap at about 1.7% in early 2026. On everyday tasks like writing, summarising, and answering questions, most people would struggle to tell the difference.
The honest caveat: benchmarks are narrow, and the largest closed models still lead on the hardest reasoning problems. The gap is small but real, and matters most at the frontier. For most practical uses, though, the choice now turns on cost, control, and customisation rather than raw ability — which is why the open ecosystem is thriving.
Try one on your own laptop
Two free tools make local AI practical on ordinary computers:
- Ollama — a lightweight command-line tool for developers. Install it on Mac, Windows, or Linux, download a model with one command, and chat in your terminal or call it from code — fully offline.
- LM Studio — a desktop app with a graphical interface: browse a catalogue of open models, download one, and start chatting — no terminal required.
The largest models still need powerful machines, but smaller ones run well on a typical 16 GB laptop — try compact versions such as Qwen3 4B, Gemma 3 4B, or Llama 3.2 3B in an efficient 4-bit format. Everything stays on your machine: private, free after download, and yours to experiment with.
What it means for you
Even if you never download a model yourself, the rise of open-source AI affects the tools you use. It drives competition, which tends to improve quality and lower prices. It gives the companies you rely on more choice, reducing the chance that a single provider controls the whole market. And it lets local languages and niche needs be served, because anyone can adapt a model rather than waiting for a big company to do it.
The story of open-source AI is still being written. What is already clear is that it has turned artificial intelligence from something a handful of companies control into something a much wider community can build with — and that shift will shape the technology for years to come. Open models are also powering a new wave of AI agents built by independent developers.
