Not long ago, the assumption in the technology industry was simple: the best artificial intelligence would belong to the few companies wealthy enough to train it, and everyone else would rent access by the token. That assumption is cracking. A wave of openly released AI models — free to download, modify and run on your own hardware — is now performing at or near the level of the leading commercial systems on many standard tests, and the consequences reach far beyond engineering.
The shift is easy to see in the release calendar we track in our technology coverage: where a frontier launch once stood alone for months, competitive open models now follow within weeks, published by research labs, startups and increasingly by large companies hedging their own bets.
Why open models caught up so fast
Three things changed. Training techniques that were once trade secrets have been published, replicated and improved in the open. The cost of computing has fallen enough that serious training runs are within reach of universities and mid-sized firms. And an ecosystem of shared tools — datasets, evaluation suites, fine-tuning frameworks hosted on platforms such as Hugging Face — means each release starts from the community’s collective work rather than from zero.
The result is a different kind of competition. Commercial labs still lead on the hardest reasoning tasks and polish, but open models win on price, privacy and control: they can run inside a company’s own data centre, be adapted to niche languages and industries, and never change behaviour overnight because a provider updated a system.
What it means for startups and users
For startups, the economics are transformative. A product that once depended on a single supplier’s pricing can now be built on a model the company hosts itself, with the freedom to switch. For users, open models are quietly appearing inside apps, creative tools and offline software that could never have justified per-query fees. Analysts following the sector, including the team at MIT Technology Review, note that the centre of gravity in AI is shifting from who owns the biggest model to who builds the most useful thing on top of widely available ones.
There are honest counter-arguments. Open releases make misuse harder to police, and the companies giving models away often do so to commoditise a layer of the stack that rivals monetise — generosity and strategy are not opposites. Quality also varies wildly, and an open model still demands real engineering to run well.
The gap that remains — and why it may not matter
Benchmarks still show the largest commercial systems ahead at the frontier. But most real-world work does not need the frontier; it needs a model that is good enough, fast, cheap and private. On that measure, the gap has effectively closed, and each month of releases — summarised regularly on our homepage — closes it further. The future of AI, it turns out, may not be a single winner’s platform. It may look more like the web itself: open foundations, built on by everyone.
The open question of openness
It is worth being precise about what “open” means here. Some releases publish full model weights that anyone can inspect and modify; others share weights under licences with conditions attached; a few merely allow free access while keeping the machinery hidden. Developers have learned to read the licence as carefully as the benchmark chart, because the freedoms that matter — commercial use, modification, redistribution — differ in ways that only appear when a project succeeds. Even so, the direction is unmistakable: each quarter, the best freely available model is better than the last, and the ecosystem of people improving it grows.
What to watch next
The next contests will be about efficiency and specialisation. Smaller open models tuned for medicine, law, code or particular languages are already beating giant general systems inside their niches, and they can run on a laptop rather than a data centre. If that trend holds, the story of AI’s second decade will not be about a handful of enormous models behind paywalls, but about millions of capable ones, owned by the people who use them. The gap that defined the first era — between those who could build AI and those who could only rent it — is the one closing fastest of all.
