Poolside's Laguna S 2.1 ships in under nine weeks

Poolside released Laguna S 2.1, a 118-billion-parameter mixture-of-experts model with 8 billion parameters active per token, on a 1-million-token context window in both thinking and non-thinking modes (Poolside blog). The timeline is the headline: Poolside's blog says the model went from kickoff to public weights in under nine weeks. In an X post, co-founder Eiso Kant shared the milestone with Latent Space's audience (Eiso Kant on X).

Poolside attributes the cycle to a "Model Factory," a programmatic pipeline that chains pre-training, post-training, reinforcement learning, and data mixtures end-to-end. Latent Space's AINews roundup frames Laguna S 2.1 as holding its own on long-horizon coding tasks against models many times its size (Latent Space AINews). A Reddit-summarized framing circulates in the same cycle: "cheaper than Deepseek v4 Flash, better than V4 Pro." That phrasing is community paraphrase, not a Poolside-issued claim.

Poolside also published the full evaluation trajectories for every benchmark trial, so the release is independently rerunnable. What remains unknown: whether outside labs reproduce the long-horizon coding results from the published trajectories, and whether the sub-nine-week cycle holds on a second model.