Laguna S 2.1

Poolside's 118B coding model with 8B active, a 1M context and the OpenMDW-1.1 license. A 96 GB Q4 file suits a 128 GB machine.

In short. Laguna S 2.1 is an open model from the San Francisco company Poolside, built for coding agents that work for hours on one problem. It has 118 billion parameters but uses only 8 billion for each token, and it can read up to a million tokens at a time. The license has almost no conditions, and a 4-bit file of 96 GB fits a Mac or mini PC with 128 GB of memory.

Summary

Poolside published Laguna S 2.1 on 21 July 2026 under the OpenMDW-1.1 license, which asks only that you keep the license text and notices when you redistribute and puts no restrictions on outputs. It is a 118B-parameter mixture of experts that activates about 8B parameters per token: 256 routed experts, 10 of them chosen per token, plus one shared expert, in 48 layers of which 12 use global attention and 36 use sliding-window attention over 512 tokens. The context is 1,048,576 tokens in both thinking and no-thinking modes. Poolside says S 2.1 scales up its XS family on the same pre-training data as XS 2.1, with a much larger post-training effort, that training took under nine weeks from start to launch, and that it was its first model trained with reinforcement learning in FP8. There are two thinking modes, off and max, with max on by default and no adjustable effort level yet. A small DFlash draft model of about 2.2 GB is published for speculative decoding.

In Poolside's own table, dated 21 July 2026 and using the highest vendor, benchmark-author or third-party score for each model, Laguna S 2.1 scores 70.2% on Terminal-Bench 2.1, 78.5% on SWE-bench Multilingual, 59.4% on SWE-Bench Pro, 40.4% on DeepSWE, 46.2% on SWE Atlas codebase Q&A and 49.7% on Toolathlon Verified. In the same table Tencent Hy3 (295B) scores 71.7%, 75.8% and 57.9% on the first three, DeepSeek-V4-Pro Max (1.6T) scores 64.0%, 76.2% and 55.4%, and Qwen 3.7 Max scores 74.5%, 78.3% and 60.6%, while Kimi K3 (88.3%) and Claude Fable 5 (88%) are far ahead on Terminal-Bench 2.1. Poolside says thinking lifts Terminal-Bench 2.1 from 60.4% to 70.2% and DeepSWE from 16.5% to 40.4%, and it publishes full trajectories for every trial at trajectories.poolside.ai. The blog lists known limitations: the model can follow its own tool format instead of a third-party harness's schema, it may write invalid JSON for tool arguments that are arrays, and it can think for a very long time before making progress.

The BF16 weights are 235.1 GB. Poolside also publishes FP8 (131.3 GB), NVFP4 (99.7 GB) and INT4 (99.7 GB) checkpoints and GGUF files of 96.0 GB at Q4_K_M, 128.8 GB at Q8_0 and 235.2 GB at F16. By our estimate the Q4_K_M file needs about 106 GB, which suits a 128 GB unified-memory machine. The card's vLLM example uses four-way tensor parallelism, and the model also runs in SGLang, Transformers, TRT-LLM and Ollama. For llama.cpp the card, last updated on 19 August 2026, points to Poolside's own fork on the branch laguna and says upstream support was still in review. In the 30 days to 11 November 2026, Hugging Face counted about 540,000 downloads of the main repository and about 223,000 of the GGUF files.

Good for

  • Coding and terminal agents that run for a long time and need a 1M-token context
  • Commercial products, since OpenMDW-1.1 asks only that you keep the license text and notices
  • Private servers or a 128 GB Mac or mini PC, using the 96 GB Q4_K_M file
  • Research on long-horizon agents, since Poolside published full trajectories for its benchmark runs

Watch out for

  • All scores come from Poolside's table, some rival scores are from third-party leaderboards, and Kimi K3 and Claude Fable 5 lead on Terminal-Bench 2.1
  • Poolside lists tool-schema drift in third-party agent harnesses, invalid JSON for array arguments and very long thinking as known problems
  • Thinking is only off or max, with no effort setting yet
  • The model card also asks that you use it consistently with Poolside's Acceptable Use Policy, which is separate from the license
  • A 1M-token context needs far more memory than the weights alone, and llama.cpp needed Poolside's fork when the card was last updated

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Estimates for the Laguna-S-2.1 (BF16) 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-S-2.1 (BF16) BF16 118B 8B active 270 GB 135 GB 81 GB Workstation
Laguna-S-2.1-FP8 FP8 118B 8B active 270 GB 135 GB 81 GB Workstation
Laguna-S-2.1-NVFP4 NVFP4 118B 8B active 270 GB 135 GB 110 GB Big-memory desktop
Laguna-S-2.1-INT4 INT4 118B 8B active 270 GB 135 GB 110 GB Big-memory desktop
GGUF Q4_K_M INT4 118B 8B active 270 GB 135 GB 106 GB Big-memory desktop

Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.

VRAM calculator

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.

More about OpenMDW-1.1

See also

Frequently asked questions

What is Laguna S 2.1?
Laguna S 2.1 is an open model from the San Francisco company Poolside, built for coding agents that work for hours on one problem. It has 118 billion parameters but uses only 8 billion for each token, and it can read up to a million tokens at a time. The license has almost no conditions, and a 4-bit file of 96 GB fits a Mac or mini PC with 128 GB of memory.
Can I use Laguna S 2.1 commercially?
Yes. Laguna S 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 S 2.1 open source?
Laguna S 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 S 2.1 on my own computer?
Only on a workstation. The smallest variant, Laguna-S-2.1 (BF16), needs about 81 GB of memory at 4-bit precision, so plan for a 48 to 96 GB GPU, several graphics cards, or a Mac with plenty of memory.
When was Laguna S 2.1 released?
Laguna S 2.1 was released by Poolside on July 21, 2026.
How long a context window does Laguna S 2.1 have?
Laguna S 2.1 supports a context window of up to 1M tokens.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
Laguna XS 2.1 Poolside 33.4B 3B active 256K OpenMDW-1.1 (Permissive) 22 GB Gaming GPU
Laguna M.1 Poolside 226B 23.4B active 256K Apache 2.0 (Permissive) 153 GB Big-memory desktop
Hy3 Tencent 295B 21B active 256K Apache 2.0 (Permissive) 204 GB Big-memory desktop
Nemotron 3 Ultra NVIDIA 550B 55B active 1M OpenMDW-1.1 (Permissive) 379 GB Big-memory desktop