Laguna XS 2.1
Poolside's 33B coding model with 3B active and a 256K context, small enough for a 36 GB Mac. OpenMDW-1.1 license.
In short. Laguna XS 2.1 is the small open coding model from Poolside. It has 33 billion parameters but uses only 3 billion for each token, so it is fast, and the model card says it runs on a Mac with 36 GB of memory. It is built for coding agents, and it is the most downloaded model in Poolside's 2.1 line.
Summary
Poolside announced Laguna XS 2.1 on 2 July 2026 as an upgraded version of Laguna XS.2, this time under the OpenMDW-1.1 license instead of Apache 2.0. It is a 33B-parameter mixture of experts that activates about 3B parameters per token, with 256 experts plus one shared expert, 40 layers of which 10 use global attention and 30 use sliding-window attention over 512 tokens, a 256K-token context, an FP8 KV cache, and interleaved thinking between tool calls that you can turn on or off per request. Poolside says the update adds 5.4 points on SWE-bench Multilingual and gives stronger results on terminal-style tasks. A small DFlash draft model is published for speculative decoding.
In Poolside's own table, Laguna XS 2.1 scores 70.9% on SWE-bench Verified, 63.1% on SWE-bench Multilingual, 47.6% on SWE-Bench Pro and 37.5% on Terminal-Bench 2.0, against 69.9%, 57.7%, 46.3% and 35.7% for XS.2. In the same table Qwen3.6-35B-A3B scores higher on all four, at 73.4%, 67.2%, 49.5% and 51.5%, and so does the 137B MAI-Code-1-Flash, at 71.6%, 65.5%, 51.2% and 54.8%. XS 2.1 is ahead of Claude Haiku 4.5 on SWE-Bench Pro and Terminal-Bench 2.0 but behind it on SWE-bench Verified (73.3%). The scores are Poolside's own, run in its agent harness with up to 500 steps and patched task images, and the comparison scores are the highest it could find published for each model.
The BF16 weights are 66.9 GB, with FP8 at 35.3 GB, NVFP4 at 21.6 GB and INT4 at 23.9 GB, and Poolside publishes official GGUF files of 20.3 GB at Q4_K_M and 66.9 GB at BF16. The card says the model is compact enough for a Mac with 36 GB of RAM, and it runs in Ollama, vLLM, SGLang, Transformers and TRT-LLM. The card, last updated on 19 August 2026, also warned about rough edges: llama.cpp support had to be built from an upstream pull request, vLLM needed a recent build for tool calls to parse correctly, and ollama run could return empty output on Apple Silicon until a Metal fix landed. In the 30 days to 11 November 2026, Hugging Face counted about 654,000 downloads of the main repository and about 112,000 of the GGUF files.
Good for
- A local coding agent on a Mac with 36 GB of memory, using the 20.3 GB Q4_K_M file
- Fast agent loops, since only about 3B parameters work on each token
- Commercial products, since OpenMDW-1.1 asks only that you keep the license text and notices
- Teams that want to fine-tune or serve a small coding model of their own
Watch out for
- Qwen3.6-35B-A3B scores higher on all four benchmarks in Poolside's own table, and all scores are Poolside's own
- The license changed from Apache 2.0 (XS.2) to OpenMDW-1.1, which is permissive but newer and not OSI-approved
- The card warned about llama.cpp needing an upstream pull request, vLLM tool-call fixes and empty Ollama output on Apple Silicon, so check the current state before relying on a Mac setup
- The card also asks that you use the model consistently with Poolside's Acceptable Use Policy, which is separate from the license
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Estimates for the GGUF Q4_K_M variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.
| Variants | Parameters | BF16 | FP8 / INT8 | 4-bit (Q4) | Fits on |
|---|---|---|---|---|---|
| Laguna-XS-2.1 (BF16) BF16 | 33.4B 3B active | 77 GB | 38 GB | 23 GB | Gaming GPU |
| Laguna-XS-2.1-FP8 FP8 | 33.4B 3B active | 77 GB | 38 GB | 23 GB | Gaming GPU |
| Laguna-XS-2.1-NVFP4 NVFP4 | 33.4B 3B active | 77 GB | 38 GB | 24 GB | Gaming GPU |
| Laguna-XS-2.1-INT4 INT4 | 33.4B 3B active | 77 GB | 38 GB | 26 GB | Gaming GPU |
| GGUF Q4_K_M INT4 | 33.4B 3B active | 77 GB | 38 GB | 22 GB | Gaming GPU |
Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.
The license in plain words
A short permissive license written for machine-learning models by the Linux Foundation and PyTorch Foundation. It covers the model, its data, documentation and software together, and puts no restrictions on what you do with the model's outputs.
See also
Frequently asked questions
- What is Laguna XS 2.1?
- Laguna XS 2.1 is the small open coding model from Poolside. It has 33 billion parameters but uses only 3 billion for each token, so it is fast, and the model card says it runs on a Mac with 36 GB of memory. It is built for coding agents, and it is the most downloaded model in Poolside's 2.1 line.
- Can I use Laguna XS 2.1 commercially?
- Yes. Laguna XS 2.1 is released under OpenMDW-1.1, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is Laguna XS 2.1 open source?
- Laguna XS 2.1 is open-weight, not open source in the strict sense. You can download the model files, but the OpenMDW-1.1 adds conditions that the Open Source Initiative definition does not allow.
- Can I run Laguna XS 2.1 on my own computer?
- Yes, on a good graphics card. The smallest variant, GGUF Q4_K_M, needs about 22 GB of memory at 4-bit precision, which suits a 24 to 32 GB gaming GPU or a Mac with 32 GB or more.
- When was Laguna XS 2.1 released?
- Laguna XS 2.1 was released by Poolside on July 2, 2026.
- How long a context window does Laguna XS 2.1 have?
- Laguna XS 2.1 supports a context window of up to 256K tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- Laguna-XS-2.1 model card (Hugging Face) huggingface.co
- Laguna-XS-2.1 license file (OpenMDW-1.1) huggingface.co
- Laguna-XS-2.1-GGUF repository (Hugging Face) huggingface.co
- Laguna M.1 and XS.2 technical report (Poolside) poolside.ai
- Poolside blog index, listing the Laguna XS 2.1 post of 2 July 2026 poolside.ai
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Laguna S 2.1 Poolside | 118B 8B active | 1M | OpenMDW-1.1 (Permissive) | 81 GB Workstation | |
| Laguna XS.2 Poolside | 33.4B 3B active | 256K | Apache 2.0 (Permissive) | 23 GB Gaming GPU | |
| Qwen3.6 Alibaba Qwen | 27.8B – 36B 3B active | 256K | Apache 2.0 (Permissive) | 19 GB Gaming GPU | |
| gpt-oss OpenAI | 21B – 117B 5.1B active | 128K | Apache 2.0 (Permissive) | 15 GB Laptop or phone |