Laguna XS 2.1: New Free Local Coding Model Tested (2026)
Laguna XS 2.1, a brand-new free local coding model from Poolside, just launched — I tested it the same day. Here’s the full breakdown.
It’s lightweight, open-source, and I’ve already got it running inside my Agent OS.
Key takeaways
Laguna XS 2.1 is a new free, open-source local coding model from Poolside — lightweight, fast, and available on Hugging Face.
In my tests it beat North Mini Code and Claude Haiku 4.5 on SWE-bench Pro, and crushed GPT-OSS — but it’s still roughly a year behind frontier models on UI polish.
You can run it three ways: inside Agent OS’s local engine, via free Claude Code, or as a Hermes agent profile — plus a free OpenRouter API if your hardware can’t run it locally.
What Is Laguna XS 2.1?
Laguna XS 2.1 is a brand-new local agentic coding model from a company called Poolside. It literally dropped less than 24 hours before I tested it, and it’s already on Hugging Face.
It’s designed specifically for agentic coding and terminal work, with a 256K context window — genuinely solid for a local model. It’s also lightweight, so it doesn’t need a monster machine to run.
How It Benchmarks
Side by side, Laguna XS 2.1 is a clear step up from its predecessor and sits close to Qwen 3.6 on SWE-bench Verified. It outperforms North Mini Code, which I’d already tested and rated as genuinely decent.
Against Claude Haiku 4.5, it comes out ahead on SWE-bench Pro — and it beats GPT-OSS by a wide margin. It’s obviously not competing with frontier models like Opus or Fable 5, and I wouldn’t expect a local model to.
What I Actually Built With It
Benchmarks only tell you so much, so I built three real things with it: a landing page, a to-do list app, and a second simple page.
Landing page — clean and simple, actually turned out better than the same test with Gemma 4
To-do app — fully functional (add, delete, mark complete), though the UI is fairly basic out of the box
Second page — solid, roughly comparable to what Claude could build about a year ago
The honest pattern with local models: they’re usually about a year behind frontier on polish, but the code that gets generated genuinely works.
My Hardware Setup
I ran this on a Mac Studio M4 Max with 36GB of memory, and it performed smoothly. One thing worth flagging: it’s not good for 3D work, so don’t expect it to handle anything in that direction.
3 Ways To Actually Use Laguna XS 2.1
There isn’t just one way to run this. I use it three different ways depending on the task:
Agent OS local engine — plugged directly into my Agent OS so everything I build lands in my workspace
Free Claude Code — swap the model Claude Code uses over to Laguna XS 2.1
A dedicated Hermes agent profile — run it agentically through Hermes alongside your other models
Because it runs locally, everything stays on your machine — nothing goes to the cloud. That means it works completely offline (handy on a plane with no Wi-Fi), and your data never leaves your device.
If your setup isn’t great for running models locally, Laguna XS 2.1 is also available for free on OpenRouter, which just dropped the same day. That means you can use it as a free API instead — no local hardware required.
How To Set It Up With Hermes (OpenRouter)
If you want to run it through Hermes via the free OpenRouter API, the setup is quick:
In your terminal, type hermes model
Select OpenRouter from the provider list
Give it a name and paste in your free OpenRouter API key
Start using Laguna XS 2.1 through Hermes, completely free
Should You Use It?
For building landing pages and simple mini apps, Laguna XS 2.1 does the job well, and it’s genuinely one of the better local coding models I’ve tested recently — better than North Mini Code, better than GPT-OSS, and close to Qwen 3.6.
Don’t expect frontier-level UI polish out of the box; you’ll either need to lower your expectations there or train it with a solid skill for design. But for free, fast, offline, agentic coding, it’s a strong addition to the local model lineup.
Where I Run All My Models
I’m plugging Laguna XS 2.1 into Goldie Bench to test it against 42 different builds alongside newer versions of Gemma and Mox — the same benchmark I use for every local model, including GLM 5.2. If you want the full Agent OS setup — Hermes, Claude Code, local engine, memory and everything else wired together — it’s inside my AI Profit Boardroom. New to this? Start free with my AI Money Lab.
