Laguna XS.2

What Laguna XS.2 is

Laguna XS.2 is Poolside's first open-weight model, meaning the trained parameters themselves are published rather than kept behind an API. It has 33 billion total parameters, of which 3 billion are active for each token it processes ^[1].

The two numbers measure different things. Total parameters describe how much the model stores; active parameters describe how much computation a single token costs. A model that only activates a slice of its weights per token can be served more cheaply than a dense model of the same total size, so the active count is the figure that matters for inference cost, while the total is the figure that matters for memory and capacity.

The weights are published under Apache 2.0 ^[1], a permissive licence that allows use, modification and redistribution.

Where it came from

The weights appeared on Hugging Face on 23 April 2026, five days before the announcement post ^[2]. The model was trained on 30 trillion tokens ^[1].

Benchmark results

Poolside reports four coding benchmarks. On SWE-bench Verified, which asks a model to resolve issues taken from real software repositories by producing patches that pass the projects' tests, Laguna XS.2 resolves 68.2 percent of tasks, against 52.0 percent for the 31B dense Gemma 4 ^[1]. On SWE-Bench Pro it scores 44.5 percent ^[1], on Terminal-Bench 2.0, which sets tasks inside a terminal environment, 30.1 percent ^[1], and on SWE-bench Multilingual, the same kind of task across more programming languages, 62.4 percent ^[1].

The comparison with Gemma 4 is the one the release leans on: a model with 33 billion total parameters but 3 billion active per token resolving more tasks than a 31 billion parameter dense model.

How the numbers were produced

Agentic coding scores depend heavily on the scaffolding around the model, so the release spells out the setup. Benchmarks ran through the Laude Institute's Harbor framework with Poolside's own agent harness, capped at 500 steps ^[1]. Sandboxes used 8 GB of RAM and 2 CPUs, while Terminal-Bench 2.0 used 48 GB and 32 CPUs ^[1]. Sampling used temperature 0.7 and top-k 20 across all runs ^[1].

These details matter because a step cap, a harness and a machine specification each change what a model can finish: more steps and more memory let an agent recover from failed attempts and run heavier tooling. Publishing them makes the numbers reproducible and lets readers judge whether a comparison to another model was run under similar conditions.

The May 2026 update

A May 2026 update moved serving to 256K context and raised the SWE-bench Verified, SWE-Bench Pro and Terminal-Bench 2.0 scores ^[1]. The longer context window determines how much code and conversation a single session can hold before earlier material has to be dropped.

Where it sits

For Poolside, Laguna XS.2 is the point at which the company began publishing weights rather than only serving a model. Apache 2.0 licensing means teams can download it, run it on their own hardware, inspect it and adapt it, which is a different proposition from a hosted coding assistant. The combination of a large total parameter count with a small active count is aimed at the trade-off that matters for self-hosting: a model that needs the storage of a 33 billion parameter system but the per-token compute of a much smaller one.