Poolside, a prominent developer in the field of specialized artificial intelligence, has officially announced the release of Laguna S 2.1, a 118-billion-parameter open-weight model specifically engineered for agentic coding tasks. This latest iteration represents a significant leap in the company’s Laguna model lineage, utilizing a Mixture-of-Experts (MoE) architecture that optimizes performance by activating only 8 billion parameters per token. Designed to bridge the gap between mid-sized efficiency and frontier-level capability, Laguna S 2.1 supports an expansive context window of up to 1 million tokens and introduces a sophisticated "thinking" mode that allows the model to allocate test-time compute dynamically. The weights for the model have been made available on Hugging Face under the OpenMDW-1.1 license, signaling a major contribution to the open-source AI community and providing developers with a high-performance alternative to closed-source proprietary systems.
Architectural Innovation and Technical Specifications
At the core of Laguna S 2.1 is its sparse Mixture-of-Experts (MoE) framework. While the model maintains a massive 118-billion-parameter footprint in resident memory, its routing mechanism ensures that only approximately 6.8% of those parameters—roughly 8 billion—are activated during any single processing step. This architectural choice is central to the model’s value proposition: it delivers the nuanced reasoning and vast knowledge base of a 100B+ parameter model while maintaining the inference speed and operational cost-efficiency typical of much smaller models.
The model is a direct scale-up of Poolside’s Laguna XS family and was trained on the same foundational pre-training dataset as the XS 2.1 version. However, the scale of Laguna S 2.1 allows for a deeper level of reasoning and a more robust understanding of complex codebase structures. One of the model’s standout technical features is its 1-million-token context window, which remains functional in both standard and "thinking" modes. This capacity allows the model to ingest entire repositories, extensive documentation, and long-running execution logs simultaneously, making it uniquely suited for "long-horizon" coding tasks where context fragmentation often leads to errors in smaller models.
Poolside has prioritized accessibility by publishing the model weights in various formats, including BF16, FP8, INT4, and NVFP4. Furthermore, official conversions for GGUF and MLX, along with DFlash draft models, have been released to ensure compatibility across a wide range of hardware and software environments.
Development Timeline and Training Infrastructure
The development of Laguna S 2.1 was characterized by an aggressive timeline and a massive allocation of specialized hardware. The training process began on May 22, 2026, utilizing a cluster of 4,096 NVIDIA H200 GPUs. From the initiation of pre-training to the final public launch, the project spanned less than nine weeks—a remarkably short window for a model of this complexity.
A significant technical milestone in the development of Laguna S 2.1 was the implementation of reinforcement learning (RL) entirely in FP8 precision. This marks the first time Poolside has utilized this precision level for the RL phase of training, a move that likely contributed to the model’s rapid development cycle and its high degree of computational efficiency. The reliance on NVIDIA’s H200 architecture allowed the team to leverage high-bandwidth memory (HBM3e), which is critical for the rapid data movement required by MoE models during both training and inference.
Benchmarking Performance: A New Leader in Open Weights
In standardized evaluations, Laguna S 2.1 has demonstrated performance that rivals or exceeds models with significantly higher parameter counts. On Terminal-Bench 2.1, a rigorous test for command-line and system-level task execution, the model achieved a score of 70.2% with its "thinking" mode enabled. This performance places it at the top of the leaderboard for open, disclosed-size models, trailing only behind much larger or closed-source frontier systems.
The model’s capabilities are particularly evident in the SWE-Bench Multilingual evaluation, where it scored 78.5%. This result effectively tops the published table for the benchmark, showcasing its proficiency across diverse programming languages and software engineering environments. In the SWE-Bench Pro (Public) benchmark, Laguna S 2.1 scored 59.4%, outperforming competitors like Tencent Hy3 (295B) and DeepSeek-V4-Pro-Max (1.6T).
Perhaps the most telling metric of the model’s efficiency is found in the DeepSWE v1.1 benchmark. Laguna S 2.1 scored 40.4%, a staggering lead over DeepSeek-V4-Pro-Max’s 9.0%, despite the latter possessing nearly six times the number of active parameters. While closed-source frontier models such as Claude Fable 5 and Kimi K3 still maintain a lead in specific high-end reasoning categories, Poolside’s data suggests that Laguna S 2.1 is arguably the most capable model in its specific weight class for autonomous software engineering.
The Paradigm of "Thinking" Modes and Test-Time Compute
A defining feature of Laguna S 2.1 is its dual-mode operation: "off" and "max." By default, the model operates in "max" mode, which enables an internal reasoning process akin to a "chain-of-thought" but optimized for agentic action. In this mode, the model does not simply predict the next token in a sequence; it sets its own test-time compute budget, allowing it to "deliberate" on complex problems before providing a final output or taking an action in a sandbox environment.
The impact of this thinking mode is quantifiable. On Terminal-Bench 2.1, the transition from standard mode to "max" thinking mode saw scores jump from 60.4% to 70.2%. Even more dramatic was the improvement on the DeepSWE benchmark, which rose from 16.5% to 40.4%. However, this increased accuracy comes at the cost of higher token consumption. For instance, DeepSWE trajectories using the thinking mode required approximately 249,000 completion tokens, compared to just 99,000 in standard mode. Poolside reports that the model is capable of maintaining coherent, productive reasoning for hours, sometimes spanning hundreds of thousands of tokens without losing focus on the primary objective.
Real-World Application: Trajectories and Case Studies
To move beyond synthetic benchmarks, Poolside released three unedited "trajectories" that illustrate the model’s behavior in real-world scenarios. These examples provide insight into how Laguna S 2.1 handles autonomous decision-making and error correction.
- Autonomous Engine Construction: In one notable demonstration, the model was tasked with building a working HTML/CSS browser engine from an entirely empty directory. Over the course of 181 steps and a 50-minute session, the model successfully constructed the engine and validated its own output against a headless Chromium instance, all without human intervention.
- Infrastructure Optimization: The model was also applied to Poolside’s own internal agent harness. It identified inefficiencies in string concatenation, replacing O(n²) operations with more efficient buffers. This resulted in the harness becoming 5.2% faster while reducing memory allocation by 71%.
- Mathematical Problem Solving: In a display of cross-disciplinary reasoning, the model re-derived Erdős Problem #397. Operating in a sandbox environment that lacked Python, the model pivoted to Perl to complete the task over a 68-minute period. Because the model’s knowledge cutoff is November 2025 and the problem was famously solved by GPT-5.2 Pro in January 2026, Poolside asserts this was an independent rediscovery rather than a retrieval of training data.
Deployment, Hardware Requirements, and Commercial Access
Despite its 118-billion-parameter size, Laguna S 2.1 is designed to be deployable on modern workstation hardware. Because every expert in the MoE architecture must remain in memory to be available for routing, the hardware requirements are determined by the full 118B count rather than the 8B active count.
At 4-bit quantization (NVFP4 or INT4), the model requires approximately 59 GB of VRAM, allowing it to fit comfortably on a single NVIDIA DGX Spark, which features 128 GB of unified memory. At FP8 precision, the requirement doubles to 118 GB, which still fits within a single DGX Spark or a single H200 GPU. For those requiring the highest fidelity at BF16 precision, the model requires 236 GB, necessitating a multi-GPU setup or two linked DGX Spark units.
Poolside has collaborated closely with NVIDIA to optimize the model for inference using the TRT-LLM framework, specifically targeting the NVFP4 format on Blackwell architecture. Day-one support has also been confirmed for popular inference engines including vLLM, SGLang, and Ollama.
For users seeking managed access, Laguna S 2.1 is available via OpenRouter. The service offers a free tier for context windows up to 256K, while the full 1M context window is available on a paid basis. The pricing structure is set at $0.10 per 1 million input tokens, $0.20 per 1 million output tokens, and $0.01 per 1 million cache-read tokens. Additionally, the model has been integrated into several AI infrastructure platforms, including Baseten, Kilo, Prime Intellect, and ZML.
Industry Implications and Future Outlook
The release of Laguna S 2.1 underscores a shifting trend in the AI industry toward specialized, agentic models that prioritize functional utility over general-purpose conversation. By providing a high-performance, open-weight model that excels at complex software engineering, Poolside is challenging the dominance of closed-ecosystem providers.
The success of the 8B-active MoE architecture suggests that "sparsity" remains the most viable path for scaling model intelligence without reaching a ceiling in hardware accessibility. As agentic AI becomes more integrated into professional software development lifecycles, the ability for a model to "think" for extended periods—as Laguna S 2.1 does—will likely become the standard requirement for autonomous systems.
Furthermore, the OpenMDW-1.1 licensing of Laguna S 2.1 provides a framework for enterprise users to fine-tune and deploy the model within private environments, a critical factor for companies concerned with intellectual property and data security in their codebases. As the industry moves further into 2026, the benchmarks set by Laguna S 2.1 are expected to serve as a baseline for the next generation of coding assistants and autonomous engineers.
