The month the frontier went open

Something changed this summer, and it happened fast enough that a lot of people missed it.

In the space of about five weeks, three models arrived that pulled the open weight world level with the closed frontier. Not close to it. Level with it.

GLM-5.2 came first, in mid June. A 744 billion parameter mixture of experts from Z.ai, with roughly 40 billion parameters active per token, a one million token context window, and its full weights published under an MIT license. On the independent Artificial Analysis index it posted the highest score any open weight model had ever reached, matching one of the strongest closed models on several long coding tasks.

Then Kimi K3 landed in mid July. Moonshot shipped a 2.8 trillion parameter sparse model, the largest open weight release to date, with native vision and an always on reasoning mode. On independent testing it placed fourth among every frontier model in the world, ahead of systems that cost several times more to use, with the weights due under a permissive license within days.

And this week Qwen 3.8 appeared. Alibaba previewed a 2.4 trillion parameter multimodal model and said the weights would open soon. There are no published benchmarks yet, so I am holding my applause on the numbers. But the direction is the whole point.

Three frontier scale models. Five weeks. All of them either open or opening. That is the line we crossed.

Why this is a threshold, not just another release

For most of the last two years, the top of the field was a walled garden. A small number of closed labs set the pace, and they set the price. If you wanted the best, you rented it through an API, on their terms, with their rate limits and their roadmap.

That arrangement quietly assumed the frontier could not be caught. This summer broke the assumption. When a model you can download for free lands fourth in the world, the distance between open and closed stops being a chasm and becomes a few weeks.

The interesting part is not really about geography. It is structural. Competition came back, and it came back through weights you can hold in your hand.

Why the return of competition is good for everyone

It is tempting to read all this as a race with a winner and a loser. I read it as a tide that lifts every boat, including the ones losing the headline today.

Open weights mean no lock-in. If a provider raises prices or changes terms, you are not trapped, because the model runs on hardware you control.

Competition at the top drives prices down. The newest Chinese frontier model is priced at roughly half the cost of the closest closed equivalent, and the mid tier now costs a fraction of last year’s rates. That pressure does not stay on one vendor. Everyone has to answer it.

Open weights make research compound in the open. When the weights and the methods are public, the next improvement builds on the last one instead of sitting locked in a vault. The whole field moves faster.

And the closed labs benefit too, even if it stings this quarter. Nothing sharpens a frontier team like a free model breathing down its neck. The result, a year from now, is better systems from everyone, at lower cost, with more places to run them.

A note before the hype takes over

I want to keep my head here. Benchmarks are noisy, and “largest” is not “best”. Qwen 3.8 has not shown a single number yet. A trillion parameter count tells you very little about serving cost until you know how many of those parameters actually fire on each token. Treat every “beats model X” claim as a place to start your own testing, not a conclusion to repeat.

But strip the marketing away and the trend is hard to argue with. The best ideas in this field are increasingly things you can download, inspect, and run yourself. That simply was not true a year ago.

Something good is happening. The frontier opened up, the competition came back, and the people who gain the most are the ones building on top of all of it.

The tech moves fast. The fundamentals do not.

Akciali