Poolside и Laguna S: 118B MoE обходит триллионную open-weights модель Thinky

Poolside AI Lab has trained Laguna S — a MoE model with 118 billion parameters that, according to the company, outperforms the open-weights model Thinky, which has around 1 trillion parameters. This was revealed by Poolside co-founder and co-CEO Eiso Kant in an interview with the Latent Space podcast, where he called the company's internal infrastructure a "model factory."

What is a "model factory"

Poolside's "model factory" is a combination of a small team of top researchers and training infrastructure built for rapid model releases. According to Eiso Kant, it was this pipeline that allowed a compact team to train Laguna S, and this is not the final point but the first result of a well-tuned process.

The key idea behind the approach is not to chase size for its own sake, but to build a repeatable training process where each subsequent model costs less and ships faster than the previous one.

• Developer — Poolside AI Lab
• Model — Laguna S, MoE with 118 billion parameters
• Claimed result — outperforms the open-weights model Thinky (~1 trillion parameters)
• Speaker — Eiso Kant, co-founder and co-CEO
• Source — interview with the Latent Space podcast

What makes Laguna S interesting

Laguna S stands out for the gap in size: 118 billion parameters versus roughly 1 trillion for the open-weights model Thinky — almost nine times smaller while, as Poolside claims, delivering higher quality. This is achieved through the Mixture-of-Experts (MoE) architecture: only part of the network's "experts" are activated for each query, so the model computes faster and cheaper than a dense model of the same nominal size.

This gap is an argument in the long-running debate over what determines model quality: raw parameter count or training engineering. Poolside is betting on the latter.

Why this matters for the industry

The emergence of Laguna S shows that a small team can compete with models an order of magnitude larger. According to Latent Space, the entire "factory" was built by a compact group of researchers — without the massive headcounts that market leaders operate with.

For the industry, this calls into question the familiar formula "more GPUs and more people means a stronger model." Poolside claims that with a correctly built process, a 118-billion-parameter model beats a trillion-parameter one, meaning the bottleneck is shifting from scale to training engineering.

"And this is just the beginning," —

Eiso Kant, co-founder and co-CEO of Poolside, on the release of Laguna S in an interview with Latent Space.

What it means

Poolside claims it has learned to release competitive models with a small team, and that Laguna S is just the first in a series. If the results are confirmed by independent benchmarks, this will reinforce a trend: training efficiency and MoE architecture are starting to matter more than a model's absolute size.

Frequently Asked Questions

What is Laguna S?

Laguna S is a Poolside language model with a Mixture-of-Experts architecture and 118 billion parameters. According to the company, it outperforms the open-weights model Thinky, which has around 1 trillion parameters.

Who is Eiso Kant?

Eiso Kant is the co-founder and co-CEO of Poolside. He is the one who, in an interview with the Latent Space podcast, revealed details about the "model factory" and the training of Laguna S.