AIbase Report: The well-known startup Poolside in the AI coding field has officially released its most powerful model to date - Laguna S2.1. This model is now completely open source and freely available on the OpenCode platform, quickly sparking discussions in the developer community.

Key Features of the Model

  • Scale and Architecture: An 118B parameter Mixture-of-Experts (MoE) model, with 8B parameters activated per token.
  • Context Length: Supports up to 1M tokens, suitable for long-term software engineering and complex agent tasks.
  • Inference Mode: Supports thinking (reasoning) and no-thinking modes, which can be switched flexibly according to needs.
  • Performance Positioning: Poolside claims that its capabilities are competitive with models with much larger parameter scales, while being compact enough to run on a single NVIDIA DGX Spark.

image.png

Open Source and Free Access Accelerate Adoption

Poolside has fully opened the weights of Laguna S2.1 under the OpenMDW-1.1 license, and developers can immediately obtain the model files on Hugging Face. The OpenCode platform has also announced that the model is freely available, emphasizing "1M Context · fully open source," greatly lowering the barrier for developers to try it out.

Community feedback indicates that the model has successfully run on various hardware, including Apple M3Max (128GB memory) through a llama.cpp fork, as well as high-throughput inference in multi-machine DGX Spark environments. In terms of benchmark testing, reports mention that it achieved a score of 70.2% on Terminal-Bench2.1 and demonstrated strong agent capabilities in coding benchmarks such as DeepSWE.

Poolside's Strategic Layout

As an AI laboratory focused on agentic coding and long-term tasks, Poolside previously launched the lightweight Laguna XS series models. The release of S2.1 marks its expansion from mid-range local models to stronger flagship capabilities. This model not only performs outstandingly but also focuses on practical deployment friendliness, providing an efficient choice for local and private deployment.

AIbase Commentary