Poolside’s Laguna S 2.1 is a recently released foundation model engineered for high-end “agentic coding”—tasks involving long-horizon reasoning and code execution. With 118 billion parameters in total and 8 billion active parameters per token under its mixture-of-experts (MoE) design, Laguna S 2.1 supports a massive context window of up to one million tokens, enabling it to operate on substantial codebases or complex engineering tasks. It was developed in under nine weeks and is aimed at delivering high performance while remaining manageable in deployment.
Key Features
- Compact scale with high performance: Despite its 118B parameter size, Laguna S 2.1 often rivals models multiple times larger on benchmarks for agentic coding. For example, it scored 70.2% on Terminal-Bench 2.1 and 59.4% on SWE-Bench Pro, metrics that place it among the top-performing models in this size category.
- Thinking modes: The model operates in two modes—“thinking” (enabled by default) and “no-thinking.” Thinking mode extends computation resources during inference to tackle harder problems, significantly improving scores compared to no-thinking mode (for example, boosting its Terminal-Bench 2.1 score).
- Reliable behavior and persistence: Training revisions focused on enabling the model to keep working through partial solutions, to re-verify its outputs, to avoid prematurely declaring success, and to persist when initial approaches fail. These behavior improvements are highlighted in case studies, such as building a browser engine from scratch, optimizing its own evaluation harness, and independently deriving a novel proof to an Erdős problem.
- Open-weight licensing and deployment flexibility: Laguna S 2.1 is fully open-weight under the permissive OpenMDW-1.1 license. This allows organizations to self-host the model, modify it, and execute it locally or in private infrastructure. It supports multiple weight formats (including BF16, FP8, INT4, etc.) and inference frameworks.
Who is it for?
Laguna S 2.1 is particularly relevant to business owners, software teams, and professionals who need powerful autonomous coding agents but want full control over data, compliance, and infrastructure. Key potential users include:
- Enterprises and regulated organizations (e.g. finance, healthcare, government) needing to keep code and data on premises.
- Development teams looking for high productivity tools for automating long complex tasks, such as building toolchains, code bases, or scientific proofs.
- Technology decision-makers balancing cost vs performance: the model’s size allows for substantial capability without the full infrastructure cost of frontier multi-trillion parameter models.
Licensing & Pricing
- License: Laguna S 2.1 is released under the OpenMDW-1.1 license. This is a permissive open license that grants unrestricted rights over the model materials (weights, architecture, documentation, etc.), including commercial use, modification, and distribution, subject only to preserving copyright notices and the license text. Outputs generated by the model are not encumbered by provider restrictions.
- Pricing / Access: The model weights are available for local deployment via repositories and inference frameworks. For hosted endpoints, pricing is available via external services: for example, OpenRouter offers a dedicated 1-million-token context deployment priced per million-token input, output, and cache-read operations. A free endpoint with limited context (e.g. 256K tokens) is also offered.
Final thoughts
Laguna S 2.1 represents a practical advancement in agentic coding models, aiming to bridge the gap between powerful closed models and lightweight open systems. Its strength lies in delivering robust performance on challenging coding tasks while offering full control via open licensing, manageable infrastructure requirements, and flexible deployment options.
However, decision-makers should recognize its limitations. It can overthink or engage in long reasoning sequences without making progress in some contexts; there is no medium level of effort control—it’s either thinking or not; and while performance is strong, it does not surpass all large closed models. For organizations that can leverage its capabilities and work within those limits, Laguna S 2.1 offers a compelling option. For those needing frontier performance with polish across wider domains (e.g., vision, large multi-modal use, or fully balanced efficiency vs scale), complementing it with other models or waiting for successive versions may make sense.
Visit the official website for more.
Keep up to date with our stories on LinkedIn, Twitter, Facebook and Instagram.