Sam: Two-point-eight trillion parameters. Open weights. And they just... posted it online.
Kai: [excited] I know! I've been refreshing that download page like it owes me money.
Sam: [dry] Obviously.
Kai: This is Kai.
Sam: And I'm Sam. It's July 19th, 2026, and we've got three open-source stories today that are all a little unhinged.
Kai: Moonshot AI just dropped Kimi K3 — a 2.8-trillion-parameter open-weight MoE with a one-million-token context window.
Sam: Then xAI open-sourced its Grok Build CLI — three days after someone caught it quietly shipping entire Git repos to a bucket it controls.
Kai: And npm v12 landed and basically nuked install scripts, Git deps, and remote sources by default.
Sam: Which is either the best security news of the decade, or the reason your CI is bright red right now.
Kai: Let's start with Kimi K3. Moonshot's calling it the largest open-weight model out of China, full stop.
Sam: Largest by parameter count — let's be precise. It's 2.8 trillion, but it's mixture-of-experts, so you're not lighting all of that up per token.
Kai: Sure, but the ambition, Sam! They're openly pitching this against Opus 4.8 and OpenAI's frontier models.
Sam: Pitching is a marketing word. Benchmarks are a science word. Where are the independent evals?
Kai: [laughs] Give it a week! But the one-million-token context is real — it's right there on the weights page.
Sam: A million-token window is great until you actually try to serve it. Do you have any idea what the memory footprint on 2.8T weights looks like?
Kai: Massive. Nobody's running this on a laptop — it's a datacenter model.
Sam: Right. So the number's impressive, but who actually gets to use it?
Sam: So realistically, this is cloud providers and big labs only.
Sam: So open weights, sure — but open access? That's a different question.
This show is made with AI: the hosts’ voices are synthetic and the scripts are AI-assisted. Every story links to its original source.