Poolside released Laguna S 2.1 this week, an open-weight large language model focused on agentic coding and long-horizon tasks. The 118-billion-parameter Mixture-of-Experts system uses 8 billion active parameters per token, supports up to 1 million tokens of context, and offers distinct “thinking” and “no-thinking” modes to balance performance and cost.
The model is engineered for persistent, extended coding and research workflows, maintaining goals, using tools, self-checking outputs, and recovering from failed approaches over hours with limited human oversight. This long-running autonomy is aimed at enterprise automation and software development use cases where reliability and cost efficiency are critical.
Laguna S 2.1 is positioned as a self-hostable alternative to metered, closed APIs, matching or outperforming larger systems on benchmarks such as Terminal-Bench 2.1 and SWE-Bench Pro. Its compact design allows deployment on a single NVIDIA DGX Spark, enabling enterprises, governments, and regulated organizations to retain data sovereignty and direct control over AI infrastructure.
The model was trained via Poolside’s Model Factory platform, which automates architecture experimentation and reinforcement learning from code execution. This pipeline allows roughly five-week iteration cycles and training completion in under four weeks on 4,000 H200 GPUs, underscoring an efficiency-first approach rather than brute-force scaling.
Laguna S 2.1 is released under the OpenMDW-1.1 license with multiple checkpoints, including BF16, FP8, INT4, and NVFP4, plus official GGUF and MLX quantizations. Weights are hosted on Hugging Face and compatible with pool, vLLM, SGLang, Ollama, llama.cpp, ZML, MLX, and NVIDIA TensorRT-LLM, as well as accessible via OpenRouter and the Poolside API.
Strategically, Poolside is targeting Western enterprises and public-sector buyers seeking trusted open-weight coding intelligence, positioning Laguna S 2.1 as an alternative to leading Chinese open models and increasingly restricted U.S. closed systems. By lowering run costs per token and enabling self-hosting, the release could reshape AI vendor selection and infrastructure strategies, marking a pivotal week in Poolside’s bid to become a core supplier of enterprise-grade coding agents.


