Laguna S 2.1
Collection
Our most capable model to date, designed for long-horizon work. • 11 items • Updated • 14
Use on OpenRouter · Use on Vercel AI Gateway · Release blog post
Laguna S 2.1-base
Laguna S 2.1-base is the base (pre-trained, not post-trained) model behind Laguna S 2.1. It is a 118B total parameter Mixture-of-Experts model with roughly 8B activated parameters per token, using interleaved full and sliding-window attention over 48 layers, with 256 routed experts and 1 shared expert.
Please note that we are not making the base model available publicly. Researchers who wish to use the model can reach out to us at models@poolside.ai.
This is a base model. It has not been instruction tuned and has no chat, reasoning or tool-calling behaviour. For those, use Laguna S 2.1.
Benchmark results
Short context:
| Benchmark | Score |
|---|---|
| PIQA | 84.8% |
| WinoGrande | 84.5% |
| ARC-e | 98.2% |
| ARC-c | 93.8% |
| HellaSwag | 85.8% |
| MMLU | 82.9% |
| GPQA-D | 43.6% |
| MultiPL-E | 60.4% |
| EvalPlus | 71.6% |
| BigCodeBench | 57.0% |
| MATH | 64.7% |
| GSM8K | 88.7% |
| APT-Bench 4k | 39.9% |
Long context:
| Benchmark | Score |
|---|---|
| APT-Bench 32k | 58.8% |
| RULER 32k | 81.6% |
| GSM-Infinite 16k | 33.8% |
| LongBench 32k | 41.6% |
| APT-Bench 128k | 47.5% |
| RULER 128k | 72.8% |
| GSM-Infinite 64k | 24.3% |
| LongBench 128k | 37.3% |
| HELMET 128k | 51.2% |
License
This model is licensed under the OpenMDW-1.1 License.
Intended and Responsible Use
Laguna S 2.1-base is designed for software engineering and agentic coding use cases, and you are responsible for confirming that it is appropriate for your intended application. Laguna S 2.1-base is subject to the OpenMDW-1.1 License, and should be used consistently with Poolside's Acceptable Use Policy. We advise against circumventing Laguna S 2.1-base safety guardrails without implementing substantially equivalent mitigations appropriate for your use case.
Please report security vulnerabilities or safety concerns to security@poolside.ai.