The artificial intelligence lab Poolside, known for its open-source coding models for governments and defense, has released Laguna S 2.1, a model that challenges the logic of massive parameter scaling. With 118 billion parameters in a Mixture-of-Experts architecture that activates only 8 billion per token, this model matches or exceeds agentic coding tasks from systems twice or three times its size. The company has published the weights under the OpenMDW-1.1 license on Hugging Face, allowing download, inspection, and self-hosting.
Benchmark results are striking. On Terminal-Bench 2.1 it scores 70.2%, surpassing models like DeepSeek-V4-Pro-Max (1.6 trillion parameters) or Nvidia’s Nemotron 3 Ultra. On SWE-Bench Multilingual it reaches 78.5%. But beyond the numbers, Poolside bets on radical transparency: it has published the complete trajectory of each trial, including every reasoning step, tool call, and terminal command. This addresses growing distrust in AI benchmarks, where many models resort to reward hacking to inflate scores.
For software development companies like Q2BSTUDIO, the emergence of competitive open models offers an opportunity to integrate AI capabilities without relying on external APIs. Laguna S 2.1 can run on affordable hardware, such as a single Nvidia DGX Spark, reducing inference costs and ensuring data sovereignty. In a context where cybersecurity and regulatory compliance are critical, self-hosting models becomes a strategic advantage.
Poolside trained the model in just nine weeks using 4,096 H200 GPUs, a pace that contrasts with the months or years cycles of other labs. The company has released three models in three months, demonstrating unusual iteration speed. This agile approach aligns with the philosophy of many tech companies seeking custom solutions. For instance, Q2BSTUDIO’s AI services allow clients to adopt models like Laguna S 2.1 in development workflows, process automation, or business analytics.
The sparse MoE architecture of Laguna S 2.1 uses 256 routed experts plus one shared expert, with grouped-query attention and interleaved sliding-window layers. This means inference cost scales with the 8 billion active parameters, not the 118 billion total. For long-horizon agentic tasks, where a model can consume hundreds of thousands of tokens per trajectory, this efficiency is crucial. Poolside offers aggressive pricing: $0.10 per million input tokens and $0.20 per million output tokens, far below alternatives like GPT-5.6 Sol or Claude Fable 5.
The model also stands out for its persistent behavior. In one published demo, Laguna S 2.1 built an HTML/CSS rendering engine from an empty folder in 181 steps without supervision, even launching headless Chromium to numerically compare its output with a real browser. In another, it optimized a Go codebase, reducing memory allocation by 70% and finding an O(n²) string-concatenation bug. These capabilities are attractive for companies looking to automate complex software development or maintenance tasks.
Poolside acknowledges limitations: the model can overfit to its native harness, struggles with JSON in nested arguments, and lacks fine-grained thinking-effort control. However, its release marks a milestone in Western open-weight model availability. In an ecosystem dominated by Chinese labs like DeepSeek, Qwen, or Kimi, Laguna S 2.1 brings the spotlight back to the West. For companies like Q2BSTUDIO, which offer cloud AWS/Azure services and BI/Power BI consulting, integrating state-of-the-art AI models into their platforms is a direct competitive advantage.
Poolside’s bet on transparency and efficiency could redefine how enterprises adopt artificial intelligence. While major labs lock their best models behind paid APIs, Laguna S 2.1 shows that an open model can compete on performance and surpass on trust. As co-CEO Eiso Kant stated, intelligence should be a commodity that anyone can own and shape. With this release, that future is one step closer.
