Vince: Okay, this one actually feels like a thing people might use. Poolside shipped Laguna S 2.1, and it’s not just another benchmark trophy with a blog post haircut.
Ava: Mm-hm. The interesting part is that it’s an open-weight coding model with a real serving story, not some vague “platform” aura cloud.
Vince: Right, and the user story is pretty obvious. If you’re building agentic coding workflows, a 1M-token context and open weights is the kind of combo that changes what you even try.
Ava: I’m with you, mostly. It’s a 118B-parameter MoE, but only about 8B activate per token, so the serving math is way nicer than the headline number makes it sound.
Vince: Exactly. That’s the part I like: all the parameters stay resident, but you’re only routing a small slice each step, which is why this can feel bigger than it is.
Ava: And they’re not pretending the size is free. At four-bit, it’s around 59 gigabytes, which is why a single DGX Spark becomes a believable target instead of a joke.
Vince: That is such an Exploring Next sentence, by the way. “Believable target instead of a joke” is basically our whole brand, somehow.
Ava: Yeah, well, the brand is surviving contact with memory limits. They also shipped weights in BF16, FP8, INT4, and NVFP4, plus GGUF and MLX conversions, which is the sort of boring distribution detail that actually matters.
Vince: And they got it out fast. Under nine weeks from training start to launch, with pre-training beginning on May twenty-second on four thousand ninety-six H two hundreds.
Ava: That’s impressive, but I’m not awarding a medal for speed alone. The thing that caught me is they say this is the first Poolside model where reinforcement learning ran in FP8 precision.
Vince: Mm-hm.
Ava: That’s a real systems clue. If you can keep RL stable in FP8, that’s not just marketing paint, that’s part of how they’re squeezing the cost curve.
Vince: Okay, zooming out a bit, this lands in a funny place. You’ve got Claude Fable 5 and Kimi K3 still leading on some of these tables, while open-weight stuff keeps getting closer in the weight class that matters for builders.
Ava: Right, and that’s the actual fight right now. Not “who is smartest in the abstract,” but who is close enough, cheap enough, and controllable enough that teams will route work through it.
Vince: Back to the numbers, though. On Terminal-Bench 2.1 with thinking on, they claim seventy point two percent, and on SWE-Bench Multilingual they report seventy-eight point five, which tops their published table.
Ava: And the headline I trust most is DeepSWE v1.1. They get forty point four percent there, versus DeepSeek-V4-Pro-Max at nine point zero, and that’s with roughly one-sixth the active parameters.
Vince: That’s the sort of thing that makes a product person perk up. Not because it wins every leaderboard, but because it suggests there’s a real efficiency win hiding in the architecture.
Ava: Sure, but don’t overread it. Poolside is careful about that, at least in the article: the claim is about the weight class, not that it’s the outright best model in the world.
Vince: Yeah, fair. And the two-mode setup is kind of the whole trick: off or max, with max on by default, and no user-facing low, medium, high effort dial yet.
Ava: And the published runs are actually the most convincing part. One built a working HTML and CSS browser engine from an empty folder over one hundred eighty-one steps, validating against headless Chromium.
Vince: That’s deliciously unglamorous. I love when a model’s big moment is basically, “congratulations, you made the browser happier.”
Ava: Ha! Right. Another run optimized Poolside’s own harness, got it 5.2 percent faster, and cut memory allocation by about 71 percent by swapping out O of n-squared string concatenation for buffers.
Vince: That’s the user story, honestly. If your model can improve the thing around it, that’s not a benchmark flex. That’s actual leverage.
Ava: The third run is also nice because it’s a sanity check, not a magic trick. It re-derived Erdős Problem number 397 in Perl over 68 minutes, because the sandbox didn’t have Python.
Vince: I mean, of course it did Perl. That’s exactly the sort of weird little detail that makes the whole thing feel more real and less demo-polished.
Ava: Yeah, and the fact that the cutoff is November twenty twenty-five matters there. It’s not just regurgitating something it saw last week.
Vince: The deployment path is pretty clean too. OpenRouter has it, and they’re saying free at two hundred fifty-six K context, then paid access at the full 1M context for ten cents, twenty cents, and one cent per million input, output, and cache-read tokens.
Ava: That pricing is doing a lot of work. It makes the model feel like something teams can actually trial without committing to a whole procurement opera.
Vince: And it’s on Baseten, Kilo, Prime Intellect’s Prime Lab, and ZML too. Plus day-one support for vLLM, SGLang, and Ollama, which is the sort of thing I wish more launches understood.
Ava: Yeah, that’s the boring adoption path that matters. If I can run it locally, serve it in the usual stacks, and not rewrite my whole harness, then it’s a real candidate instead of a press release.
Vince: Exactly. Also, I cannot believe we’re eight months into Exploring Next and still getting excited about “it works in vLLM” like it’s a holiday.
Ava: Because it is, a little. That’s the unsexy part where the actual ship happens.
Vince: So, no grand mystery here. It’s a strong open-weight coding model with real context, real routes to deployment, and enough benchmark punch to make the closed frontier people keep an eye on it.
Ava: And enough honesty in the post that I don’t feel like they’re trying to sell me a hallucination. Which, frankly, is refreshing. Okay, Vince, your turn to be annoyingly optimistic.
Vince: I’ll take it. This is the kind of release that makes me think, all right, somebody turned the model picker into something actual, and not a personality quiz.
Vince: Come on, Ava, that’s a pretty good Wednesday. Let’s leave it there before we start grading browsers for character.