Poolside has released Laguna S 2.1, a 118-billion-parameter open-weight model purpose-built for agentic coding tasks, marking a significant step in making high-performance coding AI more accessible. The model, which is already available on Hugging Face under an OpenMDW-1.1 license, utilizes a Mixture-of-Experts (MoE) architecture that activates only 8 billion parameters per token. This design allows it to deliver competitive performance against models several times its size, including DeepSeek-V4-Pro-Max, NVIDIA’s Nemotron 3 Ultra, and Thinking Machines’ Inkling, while remaining small enough to run on a single NVIDIA DGX Spark workstation.
What Is the Laguna S 2.1 Open-Weight Coding Model?
Laguna S 2.1 is a scale-up of Poolside’s Laguna XS family, trained on the same pre-training data as XS 2.1 but with a significantly larger total parameter count. Its MoE architecture means that while all 118 billion parameters reside in memory, a router network activates only roughly 6.8% of them—about 8 billion—for any given token. This sparsity is the core innovation that enables the model to behave like a much larger dense model while keeping inference costs lower. It supports a context window of up to 1 million tokens in both thinking and no-thinking modes, a key feature for working on large codebases.
How Fast Was the Laguna S 2.1 Model Trained and Released?
Poolside moved with notable speed. Pre-training for Laguna S 2.1 began on May 22, 2026, utilizing a cluster of 4,096 NVIDIA H200 GPUs. The entire process from the start of training to public launch was completed in under nine weeks. It is also the first Poolside model where reinforcement learning was run entirely in FP8 precision, a technique that likely contributed to the accelerated training timeline and increased computational efficiency.
What Benchmarks Does Laguna S 2.1 Excel At?
On long-horizon coding benchmarks, the model holds its own against significantly larger competitors. On the Terminal-Bench 2.1 evaluation, it scored 70.2 percent with max thinking enabled, placing it at the top of its weight class and above models like DeepSeek-V4-Pro-Max (64.0) and Nemotron 3 Ultra (56.4). Perhaps more striking is its performance on SWE-Bench Multilingual, where it achieved a score of 78.5, the highest among all reported models, including those with 5 to 13 times more total parameters. On DeepSWE v1.1, a benchmark focused on complex, long-horizon tasks, Laguna S 2.1 scored 40.4, a substantial improvement over DeepSeek-V4-Pro-Max’s 9.0, despite using roughly one-sixth of the active parameters. The model’s “max thinking” mode is a key differentiator, with Poolside showing that it can trade additional test-time compute for higher scores—on DeepSWE, performance jumps from 17 percent to 40 percent at the cost of roughly 2.5 times more completion tokens.
What Makes the Model’s Architecture Unique for Agentic Coding?
The MoE design is central to the model’s value proposition for agentic coding, which requires an AI to autonomously navigate, debug, and modify large codebases. The 1-million-token context window allows the model to ingest an entire repository or a long debugging session in a single pass. Combined with the efficient parameter activation, this makes it feasible to run the model on local hardware for repeated, agent-driven workflows. Poolside has also published weights in multiple precisions, including BF16, FP8, INT4, and NVFP4, alongside official GGUF and MLX conversions and DFlash draft models, giving developers flexibility in deployment based on their hardware constraints.
Real-World Trajectories Demonstrate Long-Horizon Capabilities
Beyond benchmark scores, Poolside published three unedited, long-horizon trajectories that demonstrate the model’s behavior in practice. In one instance, the model built a working HTML and CSS rendering engine from a blank folder in 181 steps over a 50-minute session, validating its output against headless Chromium. In another, it self-optimized Poolside’s own agent harness, making it 5.2 percent faster while reducing memory allocation by approximately 71 percent, by replacing O(n²) string concatenation with buffers. A third trajectory showed the model independently rediscovering a solution to an Erdos problem that had remained open for over 50 years, using Perl in an environment without Python.
Who Is the Laguna S 2.1 Model For?
This release is primarily aimed at developers, researchers, and organizations that require autonomous coding agents capable of handling complex, multi-step software engineering tasks. The open-weight nature of the model under the OpenMDW-1.1 license makes it suitable for internal fine-tuning, research, and deployment in environments where data privacy or customization is paramount. The ability to run the model on a single DGX Spark at 4-bit precision lowers the hardware barrier, making it accessible for individual developers and small teams who want to run their own coding agent without relying on cloud APIs.
What to Do Now with the Laguna S 2.1 Model
Developers and teams interested in deploying autonomous coding agents should start by downloading the model weights from Poolside’s Hugging Face repository. For those with limited local hardware, the 4-bit quantized versions and GGUF conversions are the most practical entry points, as they allow the model to run on a single NVIDIA DGX Spark or equivalent workstation. Reviewing the published agent trajectories on Poolside’s site is a strong starting point for understanding the model’s behavior in real-world coding scenarios, as they provide concrete examples of how the model plans, verifies, and iterates on complex tasks without human intervention.
