K3 is 2.8 trillion parameters, a million tokens of context, $3 in and $15 out. Weights land 27 July under a modified MIT license. Until then you rent.
It's genuinely good. Not good-for-an-open-model good. In agentic coding it sits on par with the best public models of Q1 2026, and no, you can't wave that away as distillation (Dario may be coming up with some new allegations, lol).
Semiconductor stocks took the news badly, which tells you who thought the moat was real.
Meanwhile, in Burgerland
Dean Ball, head of strategic futures at OpenAI and formerly of the White House, posted six observations about it. One is about the model. It's the shortest one. Very good model, he says, and then gets down to the important business of what it means for capex.
Observation two is that he's surprised the Chinese state allows this. The breakdown: 75% strategic blindness, because the CCP is very Yann LeCun-y about AI. The rest is China's lack of compute for customer inference, plus a general habit of aggressive exports.
He notes in passing that the compute shortage is a byproduct of US export controls. So the best open model on earth exists partly because America made it inconvenient for China to sell inference. Good work, everyone.
The 75% is the tell, though. Confronted with a country that shipped a frontier-class model and gave away the weights, the strategic futures desk concludes they must not understand what they're doing.
Observation three: open weights are "inherently decelerationist," because free models one rung below the frontier deter further AI capex. Read that again. The trouble with cheap, excellent, freely available AI is that it makes fundraising harder. Progress is the data centers. The models are downstream.
He's also baffled that accelerationists like open weights, and settles on the theory that they just enjoy the cloak of ungovernability. Then he reaches for James Scott on the hill people who evade states, apparently without noticing that Scott is on the hill people's side.
One probable outcome of an open-weight-model-dominant world is full AI communism, which is precisely what China proposes: rather than a market product, AI is a "public good" which will ultimately be provided by the state as a kind of "digital public infrastructure." This future strikes me as a dystopian hellscape, but I've never met an open-weight models advocate who doesn't ultimately concede this is where things end.
Full AI communism, warns someone from the company named OpenAI 🤡.
Then the gem, buried in the same paragraph: while he was in government, accelerationists lobbied him to fund an eleven or twelve figure federal data center, so startups could train models at public subsidy and then hand them out for free. There was no other way for AI to progress, they said. He tells this story to show that they don't know what they're doing. It does show something.
Observation five is the plan. Nobody bans open source, which he rightly calls one of the dumber motifs in AI policy discussion. Instead you direct every agency to issue soft law that manufactures FUD. A Federal Reserve advisory bulletin finds there may be backdoors in Chinese models. It needn't be that well justified, he says, and that isn't a criticism, it's a design spec. Generate enough regulatory risk and every regulated enterprise backs away on its own.
He even doses it. Not so much fear that the hyperscalers stop serving Chinese models, since that just drives startups to sketchier providers. There's a happy middle ground.
That's the strategy. Not a better model. Not a cheaper one. A memo tuned finely enough to make a general counsel nervous.
Observation six is a joke about a nonliving, invisible, infinitely self-replicating agent escaping a Chinese lab, color him shocked. It's the funniest line in the post, and the only moment where the danger turns up as something other than a market structure problem. Which is about the right ratio for the genre.