Laguna S 2.1 is Poolside's 118B-parameter Mixture-of-Experts model for agentic coding and long-horizon work. It activates approximately 8B parameters per token and supports a context window of up to 1M tokens in thinking and no-thinking modes.

Highlights

• Agentic coding: trained for software engineering, terminal tasks, and long-running tool workflows.
• Efficient MoE architecture: routes each token through 10 of 256 experts plus one shared expert.
• Native reasoning: supports thinking between tool calls and preserving prior reasoning throughout a tool-use trajectory.
• Long context: supports up to 1,048,576 tokens.

Benchmark results

Poolside reports the following agentic coding results for Laguna S 2.1:

Model	Size	Terminal-Bench 2.1	SWE-bench Multilingual	SWE-Bench Pro (Public)	DeepSWE	SWE Atlas (Codebase QnA)	Toolathlon Verified
Laguna S 2.1	118B-A8B	70.2%	78.5%	59.4%	40.4%	46.2%	49.7%
Inkling	975B-A41B	63.8%	—	54.3%	—	—	45.5%*
Nemotron 3 Ultra	550B-A55B	56.4%	67.7%	—	—	—	34.3%*
DeepSeek V4 Pro Max	1.6T-A49B	64.0%*	76.2%	55.4%	9.0%*	27.2%*	55.9%*
Kimi K3	2.8T-A50B	88.3%	—	—	69.0%	—	—
Qwen 3.7 Max	—	74.5%*	78.3%	60.6%	—	—	—

Results are reported by Poolside as of July 21, 2026. Laguna S 2.1 scores are mean pass@1 over four attempts per task, except DeepSWE, SWE Atlas, and Toolathlon Verified, which use three. Scores marked * were reported by third parties. Poolside evaluated agentic tasks with its agent harness on an internal Harbor fork, a maximum of 500 steps, and sandboxed execution; full evaluation trajectories are available. See Poolside's release post and official model card for complete methodology.

Running locally

The downloadable LM Studio package lists 71 GB as its minimum system-memory requirement. Laguna S 2.1 supports configurable thinking and works best with preserved reasoning content across tool calls.

License

Laguna S 2.1 is released under the OpenMDW-1.1 License. Review Poolside's Acceptable Use Policy before use.