Poolside Laguna XS 2.1 Review: Budget Coding Model

Six cents per million input tokens. That is the headline number for Poolside Laguna XS 2.1, the coding model Poolside shipped on July 2, 2026 — and it buys you a 262,144-token context window, reasoning support, and a 63.1% score on SWE-bench Multilingual. Most coding models that score in that range cost ten to fifty times more. So the real question this Poolside Laguna XS 2.1 review has to answer is not whether it is cheap. It is whether the model is good enough that the price feels like a loophole.

Short version: for a specific kind of work, yes. Let's get into the details.

What Is Poolside Laguna XS 2.1?

Poolside has spent the last two years building models that do one thing: write and repair software. Laguna XS 2.1 is the newest release in their small-model line — a point upgrade over Laguna XS.2 from April 2026, not a ground-up rebuild. The architecture is a mixture-of-experts design with 33 billion total parameters, 256 experts across 40 layers, and roughly 3 billion parameters active per token. That last number is why it is so cheap to serve: you get the knowledge of a 33B model while paying the compute bill of a 3B one.

The company positions it for agentic coding — the loop where a model reads an issue, explores a repository, edits files, runs tests, and iterates. That framing matters. Laguna XS 2.1 is not trying to be your essay writer or your travel planner. In general chat it is unremarkable. Point it at a failing test suite, though, and it behaves like a model twice its price class.

Poolside released the weights on Hugging Face under OpenMDW-1.1, a fully permissive license built for models and their artifacts. If you want the background, the official Poolside announcement covers the training details, and the Hugging Face model card has the deployment recipes for vLLM, llama.cpp, and Ollama.

How Good Are the Benchmarks Really?

Two numbers carry this release. SWE-bench Multilingual jumped from 57.7% to 63.1% — a 5.4-point gain in one point release, and the benchmark that best reflects real-world polyglot repositories. SWE-bench Verified moved more modestly, from 69.9% to 70.9%.

Context makes those scores interesting. A 70.9% SWE-bench Verified result puts this 33B budget model within shouting distance of flagship coding models from a year ago, and the multilingual gain suggests Poolside targeted the exact weakness small coding models usually have: everything that is not Python. If your codebase is TypeScript, Go, or Java, the 2.1 upgrade is the one that matters.

The honest caveat: benchmark deltas of one point are noise, and vendor-reported numbers deserve a raised eyebrow until independent runs land. The 5.4-point multilingual jump is large enough to take seriously. The rest, treat as directional. The fastest way to find out how it behaves on your actual code is to run your own prompt against it and a competitor side by side in CoreAI Compare — two minutes of testing beats any leaderboard.

What Does Laguna XS 2.1 Cost to Run?

Here is where the model makes its argument. These are the per-million-token rates for Laguna XS 2.1 and the models you would realistically weigh it against, as listed on CoreAI's model catalog:

Read that first row again. Laguna XS 2.1 undercuts the nearest budget competitor by more than half on input and 5x on output. Against standard-tier coding models it is roughly a twelfth to a thirtieth of the cost. An agentic session that chews through 2 million input tokens and 200,000 output tokens — a normal afternoon for a busy coding agent — costs about $0.14 on Laguna XS 2.1. The same session on Kimi K2.7 Code runs about $2.20. Cached input drops to $0.03 per million, which agentic loops hit constantly since they re-read the same files.

There is one structural limit hiding in the spec sheet: output is capped at 32,768 tokens per response. For iterative agent loops that edit a few files per step, that is fine. For one-shot "generate the whole module" prompts, larger models with six-figure output caps have the edge.

When Should You Pick It Over a Bigger Coding Model?

Pick Laguna XS 2.1 when the work is high-volume and verifiable. Test-driven bug fixing, dependency upgrades, lint sweeps, writing unit tests, translating code between languages, batch refactors across a large repo — anywhere the loop runs many times and a test suite catches mistakes, the cost advantage compounds and the occasional miss costs you a retry, not a production incident. It is also the obvious choice for learning agentic workflows, because experimenting at $0.06 per million tokens removes the fear of a runaway loop.

Skip it when the task is architectural. Designing a new system, untangling a gnarly concurrency bug, or making judgment calls across a sprawling monorepo still favors heavyweight models — that is where something like Grok 4.5 or a frontier reasoning model earns its price. A sensible setup uses both: the big model plans, Laguna XS 2.1 executes. And if budget-tier models are your whole strategy, our roundup of the best budget AI models of 2026 maps the full field.

One more practical note: because the weights are open, teams with strict data policies can self-host the exact same model they prototype with in the cloud. Not many models in this price class offer that escape hatch.

How to Try Laguna XS 2.1 Without an API Key

You do not need to sign up for a developer account or wire up billing to test it. Laguna XS 2.1 is live on CoreAI alongside 300+ other models — open the app or the web version, pick it from the model list, and start a chat. Because CoreAI is one subscription across every model, you can bounce between Laguna XS 2.1 for grunt work and a frontier model for hard problems inside the same conversation history, on iOS, Android, or the web.

The move that actually settles a "is this model good enough?" debate: take a real bug from your backlog, paste it into Compare, and run Laguna XS 2.1 against whatever you use today. If the budget model matches the expensive one on your task — and on routine coding work it often will — you have your answer. Developers who want quick utilities alongside chat can also lean on CoreAI's free AI Code Explainer for reading unfamiliar snippets without burning a chat session.

Frequently Asked Questions

Is Poolside Laguna XS 2.1 free to use?

The weights are openly available under the OpenMDW-1.1 license, so you can download and run the model yourself at no license cost. Hosted access is priced per token — on CoreAI it is a budget-tier model at $0.06 per million input tokens and $0.12 per million output tokens, among the cheapest coding models in the catalog.

How does Laguna XS 2.1 differ from Laguna XS.2?

It is a point upgrade focused on coding quality. SWE-bench Multilingual rose from 57.7% to 63.1% and SWE-bench Verified from 69.9% to 70.9%, with the multilingual gain being the standout. Architecture, context window, and the small-model footprint stay the same.

What context window does Laguna XS 2.1 support?

262,144 tokens of input context, with responses capped at 32,768 output tokens. That input window comfortably holds a mid-sized repository's relevant files plus a long agentic conversation, but the output cap means very large single-shot generations need to be split across turns.

Is Laguna XS 2.1 good for languages other than Python?

That is precisely what the 2.1 release targeted. The 5.4-point jump on SWE-bench Multilingual is the biggest improvement in this release, so TypeScript, Go, Java, and other non-Python codebases are where you will feel the upgrade most.

Can I compare Laguna XS 2.1 against other coding models before committing?

Yes. CoreAI's Compare feature runs the same prompt through multiple models side by side, so you can test Laguna XS 2.1 against Kimi K2.7 Code, KAT-Coder-Pro, or any of the 300+ models in the catalog with your own real tasks.