AI Pulse · AI Trends Pulse

Poolside just released Laguna S 2.1, a 118 billion parameter coding model that uses mixture-of-experts architecture to activate only 8 billion parameters at a time. That matters because it means you can run a genuinely capable agentic coding assistant on hardware you own, specifically a single NVIDIA DGX Spark box, without paying per-token to OpenAI or Anthropic.

The model handles up to one million tokens of context, which is enough to work across multiple files in a real codebase. It’s open-weight, so you can download it, inspect it, and deploy it behind your firewall. For teams working on proprietary code or operating in regulated industries, that changes the cost structure and the risk profile. You’re not sending your codebase snippets to a third party API every time a developer asks for help refactoring a function.

This isn’t theoretical. According to the announcement, teams can self-host this today if they have the box. The DGX Spark isn’t cheap, but for a company already spending five figures a month on API calls or worried about code leaving the building, the math starts to work. Agentic coding models that can reason across tasks, suggest architecture changes, and write tests used to require either frontier lab access or a research team. Now you can run one in your own data centre.

If you’re evaluating whether to build internal AI tooling or keep renting it, this is the kind of capability we help teams integrate into an AI command centre that sits inside your infrastructure. The shift from “only the labs can do this” to “you can host this yourself” happened faster than most procurement cycles move. Laguna S 2.1 won’t replace every use case, but it’s a credible option for teams that want control and predictable costs.