Laguna XS 2.1

Laguna XS 2.1

Writes code and works through long coding tasks with tool calls, using a 33B mixture-of-experts model that activates 3B parameters per token. For developers who run coding agents locally.

Made by poolside on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Laguna XS 2.1 is a text generation model from poolside, built for agentic coding and long-running work on a local machine. It is a mixture-of-experts model with 33B total parameters and 3B active for each token. It upgrades the earlier Laguna XS.2 and scores higher on several coding benchmarks.

Who it suits: Developers who want a coding model they can run on their own hardware, with a long context and optional step-by-step thinking.

Industry | Who uses it | How they use it | Example
Software | Developers | Run a coding agent on a local machine | Ask it to fix a bug across a repository
Open-source work | Contributors | Draft changes and tests with a permissively licensed model | Write a patch and the test that covers it
AI research | Model evaluators | Compare a small active-size model with larger ones | Check the card's benchmark table against your own tasks

Anyone who needs a model for images or audio can skip it, because the card lists text in and text out.

Where to use it:

• Python with the transformers library: the record lists it as the library.
• Ollama and llama.cpp: the card says the model is available there.
• OpenRouter: the card links a hosted version.
• A Mac with 36 GB of RAM: the card says the model is compact enough for it.

Use case | Who | What they get back
Agentic coding | Developers | Code changes produced through tool calls and short thinking steps
Long tasks | Developers | Work across a context window of 262,144 tokens
Local runs | Developers on a Mac | A quantized copy that fits a machine with 36 GB of RAM

How to set it up: What you need first:

• Python and pip, to download the files.
• Memory for the weights: about 66.9 GB in BF16, or about 16.7 GB at 4-bit as an estimate.
• Or a quantized copy: the card names FP8, NVFP4 and INT4 variants.

1. Install the Hugging Face hub package with pip.
2. Download the model with the hf command.
3. Load it with transformers, or run a quantized copy in Ollama or llama.cpp.

pip install -U huggingface_hub
hf download poolside/Laguna-XS-2.1

How to use it: You give it a coding task in chat form. It can think between tool calls, and you can turn thinking on or off for each request. The card's usage section is not part of the text available here, so check the model page for code.

Example: You ask it to find and fix a failing test in a project, and it has tools for reading files and running commands. It works through the task step by step, thinking between tool calls, and returns a proposed fix.

For newcomers and experts: This is a language model tuned for programming work. A good first step is to try a quantized copy in Ollama on a machine with enough memory, and give it one small coding question.

Only 3B of the 33B parameters run for each token, which keeps the model light for its size. Of its 40 layers, 10 use global attention and 30 use a 512-token sliding window. The KV cache is stored in FP8. The trade-off is that its benchmark scores sit below some larger models in the card's table.

Send a coding task with its tools
The model thinks, then calls a tool
It reads the tool result and continues
It returns the final answer or patch

Good to know: The card compares scores on coding benchmarks. This chart shows SWE-bench Verified.

Title: SWE-bench Verified, percent solved
Laguna XS 2.1: 70.9 percent
Laguna XS.2: 69.9 percent
MAI-Code-1-Flash: 71.6 percent
Claude Haiku 4.5: 73.3 percent
Qwen3.6-35B-A3B: 73.4 percent

• Size: 33.4B parameters, with 256 experts and 1 shared expert.
• Context: 262,144 tokens.
• Terminal-Bench 2.0: 37.5 percent for this model, against 35.7 percent for the earlier XS.2.
• License: OpenMDW-1.1, which the card describes as fully permissive, including commercial products.
• Cost to run: no API key or account is needed for the weights. The card names OpenRouter for hosted use and offers support on request.
• Data it reaches: files you download from Hugging Face, and the prompts and code you give it.
• Not for: images or audio input, since it is text to text.

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 33.4B
• Weights: about 66.9 GB in BF16; about 16.7 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 633,340
• Likes: 268
• License: openmdw-1.1
• Quantized versions: 39
• Fine-tunes: 13
• Tags: transformers, safetensors, laguna, text-generation, laguna-xs-2.1, vllm, conversational

Links

• Model: https://huggingface.co/poolside/Laguna-XS-2.1
• Files: https://huggingface.co/poolside/Laguna-XS-2.1/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download poolside/Laguna-XS-2.1

Then follow the usage in its model card. Nothing in this kit runs by itself.