Laguna S 2.1
Release in the Laguna family · version S 2.1
Laguna S 2.1 is a mixture-of-experts model with 118B total and about 8B active parameters per token, built for agentic coding. It has 48 layers mixing global and sliding-window attention, 256 routed experts plus one shared expert, and a 1,048,576-token context window, with reasoning that can be switched on or off per request.1
- Model hub: Model card (external site: huggingface.co)
- License: License (OpenMDW-1.1) (external site: huggingface.co)
- Release notes: Release announcement (external site: poolside.ai)
- Documentation: Evaluation trajectories (external site: trajectories.poolside.ai)
Availability and license
Overall availability
Weights download from Hugging Face without an access gate. Also served through the Poolside API and OpenRouter.17
Availability is separate from permission: read the license before using or redistributing.
OpenMDW-1.1 grants broad use rights and places no restrictions on outputs. It requires keeping the license and notices when redistributing, and ends the rights of anyone who sues claiming the materials infringe a patent or copyright. The model card also asks users to follow Poolside's Acceptable Use Policy.21
Model-disclosure tier
The model parameters for this release can be downloaded by the public. License terms may still restrict use, redistribution, or commercial use.
The weights are under a license that is not on the rubric's OSI-approved list. Read its terms before use.
Public materials checklist
| Item | Status | Notes and evidence |
|---|---|---|
| WeightsCan the general public download the model parameters for this release? | Public | BF16 weights are in the ungated repository; FP8, NVFP4, INT4, and GGUF variants are published as separate repositories.1 |
| Inference codeIs code for running the model published? | Public | The repository includes modeling code (modeling_laguna.py), and the model card documents vLLM, SGLang, TensorRT-LLM, llama.cpp, and Ollama.13 |
| Training codeIs the code used to train the model published? | Unknown | Not assessed. |
| Training-data informationPublic = the training data itself can be obtained. Partial = composition or sources are documented without full access. | Partial | The release blog says S 2.1 was pre-trained on exactly the same data as Laguna XS 2.1 and describes post-training task sources (open-source repositories, internally synthesized tasks, and tasks acquired from external data vendors) at a high level. The pre-training data composition is not described for this release, and the data is not released.4 |
| Training recipeAre the training configuration and procedure documented in enough detail to follow? | Partial | The release blog describes the training at a high level: a scale-up of the Laguna XS family with training-code fixes and small recipe changes, an SFT stage partly using synthetic data, and reinforcement learning run in FP8 precision. The blog does not give hyperparameters or training configurations.4 |
| Evaluation materialsPublic = evaluation code or prompts that let others re-run the evaluations are published. Partial = results only. | Partial | The model card publishes benchmark results, and Poolside publishes full evaluation trajectories.15 |
What it is useful for
Poolside documents it for long-running coding and research tasks with extended tool use, and for running an open-weight model locally on high-memory hardware.6
Run and use notes
- The model card states that the BF16 checkpoint is roughly 236GB of weights and needs multiple GPUs, and that quantized variants reduce this. Its vLLM and SGLang examples use tensor parallelism of 4.1
- llama.cpp support uses Poolside's llama.cpp fork (laguna branch); the model card notes that base support is under review upstream.1
Organization context
Provenance and derivatives
Other releases in the Laguna family
- Laguna XS 2.1Model-disclosure tier (USASI rubric v0.1): Open-weight
U.S. eligibility
Sources
This listing is not an endorsement, a safety assessment, or a federal approval.