Video: "Laguna XS 2.1: Poolside's Local Coding Agent Tested - Nine Languages" by Julian Goldie on YouTube.
What Laguna XS 2.1 is
Poolside is a US AI company focused on software engineering tools. Laguna XS 2.1 is their compact local coding model — 33 billion total parameters arranged as a mixture-of-experts architecture with 3 billion active parameters per inference pass. That design keeps the resource footprint small: the model runs on a mid-range GPU rather than requiring server-grade hardware.
The 2.1 release is specifically positioned as a local alternative to cloud coding models for teams that want to keep code off external APIs, or for individual developers who prefer not to pay per-token for routine coding tasks. It is not a general-purpose model; the training data and fine-tuning are focused on software engineering tasks.
What the SWE-bench Pro result actually means
SWE-bench Pro is a benchmark that tests a model's ability to resolve real GitHub issues in actual codebases — more representative of genuine development work than typical code completion benchmarks. Laguna XS 2.1 scores above both North Mini Code and Claude Haiku 4.5 on this benchmark, which is notable given its size and the fact that it runs locally.
The comparison matters because Haiku 4.5 is a commonly used lightweight cloud model for coding tasks, and North Mini Code is a dedicated small coding model. Beating both on a practical benchmark while running on your own hardware makes a genuine case for the model in cost-sensitive environments. That said, SWE-bench Pro tests structured issue resolution — the comparison does not extend to greenfield architecture decisions or complex multi-file refactors.
What Julian Goldie found across nine languages
Julian Goldie's test covered Python, JavaScript, TypeScript, Go, Rust, Java, C#, PHP, and Ruby. The results break roughly into three tiers. For Python, JavaScript, and TypeScript the output was clean and immediately usable — consistent syntax, correct structure, sensible variable naming. For Go and Rust the model produced working code in most cases but showed occasional lapses in idiomatic style that would need a review pass. For Java, C#, PHP, and Ruby performance was more variable, with the model sometimes producing valid but verbose output that a language-familiar developer would simplify.
The practical takeaway from the test is that Laguna XS 2.1 is a solid choice for teams working primarily in the web-native languages where it performs most consistently. If your stack is Go or Rust, the model is useful but not a drop-in replacement for a cloud model on those tasks. The lower-coverage languages require more review.
Running it locally: what the setup involves
Laguna XS 2.1 runs via Ollama, which means the installation process is straightforward for anyone already using local models. You pull the model, point your IDE extension or coding agent at it, and it runs inference on your own machine. There is no external API call, no token cost, and no code leaving your environment.
On a machine with a capable consumer GPU — an RTX 4080 or equivalent — response times are quick enough for interactive coding use. On CPU-only machines the model will run but latency increases to a point where it is better suited to batch tasks than interactive completion.
Where this fits in a UK development team's tooling
For development teams handling client code, sensitive internal systems, or anything subject to data handling agreements, local models eliminate the compliance question entirely. Laguna XS 2.1 offers a route to capable AI-assisted coding without routing any code through a third-party API. The benchmark results suggest it is genuinely competitive for the languages most UK web-focused teams use day to day.
The trade-off is setup effort and hardware dependency. Cloud models are available immediately from any machine; a local model requires a capable GPU and a working Ollama install. For an individual developer that is a one-hour setup. For a team it requires a shared infrastructure decision or per-machine installs.
Where this connects to NordSys
We help UK development teams choose and integrate AI coding tools — including local models for teams with data handling requirements or cost sensitivity. If you want to understand whether Laguna XS 2.1 or a comparable local model fits your stack, we can assess your setup and recommend what makes sense.
See our Programming service →