The Kimi K3 countdown hit zero this morning — and Hugging Face is still showing you a clock, not a download. This is the AI Daily Briefing. Today — a signed $10 billion Korea deal, two open-model releases dropping the same day, and one broken promise. Where's the receipt? This one comes via Brandon Fisher at TechTimes. Okay, this is the one I've been building toward all week. NAVER and NVIDIA signed a 200-megawatt, ten-billion-dollar deal, with Brookfield as the infrastructure capital partner. Brookfield's role is the detail that matters. Remember the SK-NVIDIA five-hundred-billion letter of intent we walked through Sunday? This one's signed and financed. Two-gigawatt LOIs make for a press release. Two hundred megawatts with Brookfield's checkbook attached means concrete gets poured. And it's running Vera Rubin and Blackwell silicon. So I keep coming back to one chokepoint: where does the HBM4 come from to feed 200 megawatts of Rubin? SK Hynix is right there in-country, sure. But 'right there' doesn't automatically mean 'fast enough to justify the build timeline.' NVIDIA also owns a slice of the operator. So they sell the chips and own the landlord — vertically integrated all the way down to the power bill. This one deal triples Korea's whole AI factory capacity. The dollar figure grabs the headline, but the megawatts available on a live grid are the real constraint. Ten billion buys the chips. It doesn't manufacture the memory any faster. From DEV Community:
Moonshot AI's Kimi K3 is not yet an open-weight model. The official Kimi K3 page on Hugging Face is currently an "Upcoming release" countdown pointing at July 27, and the moonshotai organization listing contains no K3 repository, no model card, and no license file.
Yesterday I closed with, 'assuming the weights actually show up tomorrow.' Well, tomorrow's here. The Hugging Face page is still a countdown timer — no repo, no model card, no license file. Back on the 26th, I said, 'Today it's an API and a leaderboard slot.' Well, here we are. It's still an API and a leaderboard slot. Moonshot's launch post calls it 'the world's first open 3T-class model.' That's 2.8 trillion parameters you can't download. Here, the word 'open' is attached to a paid endpoint. Right, and they missed a date they chose: July 27. They put a clock on it and told the open-source community to wait. The artifact they promised still isn't there. The only live product is behind a paywall. If the weights ever land, 2.8 trillion parameters with native vision and a million-token context would be a genuinely big release. Until then, it's a screenshot and a countdown. I'll get excited when there's a checkpoint I can actually pull and blow up my latency budget with. Not before. This one's from The News 92:
A small group of AI researchers (Reactor) have released Open Dreamer, an open implementation of the Dreamer 4 world-model pipeline written in JAX and Flax NNX. Two repositories were released. next-state/open-dreamer holds the training pipeline: a causal video tokenizer, an action-conditioned latent dynamics model, rollout generation, and FVD scoring.
So a small crew called Reactor just reproduced Dreamer 4 in JAX and Flax. What got me was the browser demo — a Game-to-Dream toggle that hands a live Minecraft stream over to the world model frame by frame. There's an actual technical write-up behind it, too, covering a causal video tokenizer, action-conditioned latent dynamics, and FVD scoring. And two repos back up the demo. Right. They were disciplined about it — starting on CoinRun, a 2D platformer you can train on a single GPU, then scaling the same pipeline up to Minecraft. You can actually follow the engineering from one to the other. It also quietly answers the question of whether everything has to be a Transformer forever. Latent dynamics world models are a different workload, and they're not going away. One backbone, two models: space layers inside a frame, causal time layers between frames. Clever. Before I get too excited, show me the rollout drift — how many frames until the dream falls apart? This one's from AIToolly:
Meituan's technical team has officially open-sourced LongCat-2.0, a massive model featuring 1.6 trillion total parameters with an average activation of approximately 48 billion. Specifically engineered for "Agentic Coding" tasks, the model introduces significant architectural innovations, including LongCat Sparse Attention and N-gram Embedding.
Meituan open-sourced LongCat-2.0 — 1.6 trillion parameters, 48 billion active. And everyone's gonna headline the 1.6T. But even with only 48B active, you still have to hold the full 1.6 trillion parameters in memory to serve it. That's where the cost bites. What stands out to me is the domestic inference code. They shipped specialized kernels for Chinese computing cards — non-NVIDIA silicon. Control the fine-tuning and inference stack, and you control the model. Meituan's answer is clear: not Jensen. Right, and that's a supply-chain bet dressed as a coding release. The number I want to pull apart is cost per token on domestic iron versus an H-series rack. Nobody in the press release wants to show you that math. They call it 'Agentic Coding.' A 1.6T agentic model drops with no step-level error rate and no observability story. After the week we've had, you notice that omission. Have feedback, a story idea, or a correction? Email us at aidailybriefing at lantern podcasts dot com. We’d love to hear from you.
What we’re watching next: Moonshot’s July 27 Hugging Face window. When do the Kimi K3 files arrive, along with a model card and license?
Links to every story are in today’s show notes if you want to dig into anything that caught your attention.
That’s AI Daily Briefing for today. This is a Lantern Podcast.