Sam: 8,192 tokens. [beat] That's the default amount of thinking it does before it even answers you.
Kai: I'm Kai.
Sam: And I'm Sam.
Kai: And this is Open Source After Dark. It's Monday, August 17th, and we've got a packed show for you.
Kai: On the show today: a new model from Alibaba is punching way above its weight, but it comes with a bizarre quirk.
Sam: We'll also look at Nvidia quietly walking back a massive, quarter-trillion-dollar guarantee for OpenAI.
Kai: And we'll wrap with a bombshell acquisition: Stripe is reportedly buying AI gateway OpenRouter for a number that will make your eyes water.
Kai: Alright, let's dive into this new Qwen 3.8 27B model. Simon Willison has a fantastic deep dive, and Sam, this thing is a monster.
Sam: It's putting up numbers that rival models twice its size. On benchmarks like MMLU, it’s going toe-to-toe with the heavyweights. I'll give you that, it's impressive.
Kai: [excited] See? It's a huge leap for open models in this size class! This is the kind of efficient power we've been waiting for.
Sam: Okay, but let's circle back to that cold open. Out of the box, this thing overthinks. Its default max tokens is set to 8,192.
Kai: So it's thorough. You can just change the parameter—
Sam: It's not 'thorough,' Kai, it's a performance disaster waiting to happen. It will literally keep generating text until it hits that massive limit unless you stop it. For production APIs, that's slow, expensive, and completely unpredictable.
Kai: That's a guardrail, not a mandate! The point is, the capability is there. You just tune it down. This is a model for developers!
Sam: Defaults matter. How many people are going to download this, run it with the defaults, and complain that it's unusable? It's a self-inflicted wound.
Kai: So the takeaway for developers is simple: grab this amazing model, but for God's sake, set `max_tokens` to something sane, like 512 or 1024.
Sam: And maybe set the `repeat_penalty` lower than 1.2, unless you want it to get really creative and avoid repeating itself at all costs.
Kai: Look, bottom line: it's a top-tier model that you can run on consumer hardware, with a few quirks. That's a huge win for the open-source community.
Sam: Alright, let's move on. It sounds like things are getting tense between Nvidia and OpenAI.
Kai: 'Tense' might be a strong word. Reuters is reporting Nvidia's scaling back the financing it might guarantee for OpenAI's data centers. This was always a hypothetical, remember.
Sam: The hypothetical was for up to 250 billion dollars. A quarter of a trillion. That's not just some number you throw around.
Kai: Right, and now they're just adjusting it to a more realistic level. This isn't a pull-out, it's a recalibration.
Sam: You don't 'recalibrate' from a number that big without a reason. To me, it signals that Nvidia is getting worried about the risk. Maybe the projected demand for that much compute isn't as ironclad as we thought.
Kai: Or... maybe OpenAI is just so successful they've found other financing partners and don't need Nvidia to underwrite the entire planet's supply of GPUs for them.
Sam: [dry] That's a very generous interpretation, Kai. The market is seeing this as a sign of nervousness. The era of the blank check for AI infrastructure might be ending.
Kai: Okay, so what's the takeaway for developers?
Sam: Well, it means the cost of compute might not keep falling at the subsidized rate we've all enjoyed. When the people footing the bill for data centers get nervous, that nervousness eventually trickles down to our API bills.
Kai: I still think this is just a sign of a maturing, diversifying market. Let's not panic.
Sam: Famous last words. Speaking of mature markets—
Kai: Speaking of mature markets... TechCrunch is reporting a huge move: Stripe is acquiring OpenRouter.
Sam: And the price tag... more than seven billion dollars.
Kai: [excited] Seven BILLION! It's incredible. It makes perfect sense. Stripe handles payments, OpenRouter handles AI model routing. It's the full stack for any modern application.
Sam: Hold on. OpenRouter is a smart service. It finds the best or cheapest model for your query and routes it. It's a great feature. Is it a seven billion dollar company?
Kai: It's not about what it is today, it's about what it will be. Stripe is buying the central switchboard for AI! Every developer using multiple models is either already on OpenRouter, or building their own version of it.
Sam: And I loved that it was an independent Switzerland, helping you find the truly best option. Now it's part of Stripe. What happens when the 'best' route is the one that keeps you in the Stripe ecosystem?
Kai: That's so cynical! This is about integration and developer experience! One bill for your payments and your AI. It's brilliant!
Sam: Okay, so the advice for developers is: yes, a single API could make your life easier. But you have to watch the routing like a hawk. Verify the choices it makes. The moment it stops being impartial, its value evaporates.
Kai: It's a huge bet on the future of AI embedded in every app. And Stripe just bought a pole position.
Sam: Alright, let's do a quick recap.
Kai: A powerful new Qwen model is out, but it thinks a little too hard by default.
Sam: Nvidia is getting cold feet on a quarter-trillion-dollar guarantee for OpenAI.
Kai: And Stripe is buying the entire AI switchboard by acquiring OpenRouter for over $7 billion.
Sam: And before we go... a quick note. Today, August 17th, is apparently Black Cat Appreciation Day.
Kai: [laughs] Seriously? Okay, in that spirit, shoutout to all the great cat-related repos on GitHub. And I don't just mean `cat` the command. I mean the important stuff, like ASCII art generators.
Sam: A vital part of the open-source ecosystem. Just don't ask a Qwen model to draw one with the default settings, it might take all day.
Kai: That's our show! I'm Kai.
Sam: And I'm Sam. We'll see you tomorrow.
This show is made with AI: the hosts’ voices are synthetic and the scripts are AI-assisted. Every story links to its original source.