Poolside AI releases Laguna S 2.1, a 118B-parameter MoE model outperforming larger open-weight models

Poolside AI, co-led by Eiso Kant, has released Laguna S 2.1, a 118-billion-parameter Mixture-of-Experts (MoE) model with 8 billion parameters activated per token. The company asserts that this model outperforms larger open-weight models, including a roughly 1-trillion-parameter model from Thinking Machines, which is nearly ten times its size.

To achieve this, Poolside developed a system it calls the 'Model Factory,' an end-to-end pipeline designed to rapidly train and improve models. The system supports 10,000–20,000 experiments per month, enabling the team to move from pre-training to model release in as little as eight weeks. Key components include streaming data directly into training, reproducible experimentation, and low-precision compute optimizations.

Laguna S 2.1 introduces a context window of up to 1 million tokens and offers both 'thinking' and 'no-thinking' modes. The model’s architecture and training pipeline are detailed in a technical report praised for its transparency, according to Poolside’s co-CEO.

Poolside’s approach emphasizes engineering efficiency, with Eiso Kant noting that 95% of model-building improvements can be attributed to better data or compute efficiency. The company also highlights the role of persistence, verification, and backtracking in outperforming raw model intelligence, particularly in long-horizon tasks.

The release follows Poolside’s $500 million funding round and reflects its commitment to open-weight and open-research principles. Kant argues that a competitive landscape with many foundation-model companies is preferable to an oligopoly, even if Poolside becomes one of the dominant players.