Poolside Unveils Laguna XS 2.1: A New Benchmark for Open-Weight Agentic Coding

In addition to SWE-bench, the model shows steady improvements in SWE-bench Verified, climbing from 69.9% in XS.2 to 70.9%. These metrics confirm that the model is not just memorizing patterns but is developing a more robust ability to navigate complex, multi-step coding tasks—the hallmark of a true agentic model. For developers, this means fewer hallucinations and more reliable code generation in production-grade environments.

• SWE-bench Multilingual: 63.1% (up from 57.7%)
• SWE-bench Verified: 70.9% (up from 69.9%)
• Significant gains via data and training refresh

Deployment & Quantization

Poolside has gone all-in on developer accessibility. Laguna XS 2.1 is supported across the most popular inference engines, including vLLM, SGLang, NVIDIA TensorRT-LLM, HuggingFace Transformers, and Ollama. This ecosystem support ensures that whether you are deploying on a massive cloud cluster or a local workstation, you have a path to production.

To cater to different hardware constraints, several quantized checkpoints are available immediately. Developers can choose from FP8, INT4, and NVFP4 formats to optimize for memory bandwidth and VRAM usage. For those in the llama.cpp community, GGUF support is expected to follow shortly, ensuring that even users with consumer-grade hardware can run this powerhouse locally.

• Supported Engines: vLLM, SGLang, TensorRT-LLM, Transformers, Ollama
• Available Quantizations: FP8, INT4, NVFP4
• GGUF support coming soon via llama.cpp

API Pricing

For those who prefer managed infrastructure, Laguna XS 2.1 is available via the Poolside API and OpenRouter. The pricing structure is highly competitive, specifically designed to reward efficient usage through aggressive cache-hit discounts. This makes it an excellent choice for agentic loops that frequently reference the same codebase or system prompts.

The 256K context window is fully supported on the API, allowing for deep-context reasoning without the need for complex RAG architectures in many scenarios. Note that the previous XS.2 model will be sunset on the Poolside API after one week, so developers should migrate to 2.1 immediately to take advantage of the improved performance and pricing.

• Input Price: $0.10 per 1M tokens
• Input (Cache Hit): $0.05 per 1M tokens
• Output Price: $0.20 per 1M tokens
• Context Window: 256,000 tokens

Use Cases

Laguna XS 2.1 is purpose-built for agentic coding. This means it excels in scenarios where the model must act as an autonomous agent—planning a task, writing code, running tests, and iterating based on error messages. Its ability to handle long-horizon work makes it perfect for automated refactoring, bug fixing, and even generating comprehensive unit test suites.

Beyond pure coding, its high-reasoning capabilities and large context window make it suitable for complex RAG (Retrieval-Augmented Generation) applications involving large technical manuals, or as a reasoning engine for technical documentation and architectural planning. It is the ideal balance of speed, cost, and intelligence for the modern AI engineer.

• Autonomous Coding Agents
• Automated Bug Fixing & Refactoring
• Large-scale Codebase Analysis (RAG)
• Long-horizon technical reasoning

Getting Started

Ready to integrate Laguna XS 2.1 into your workflow? You can access the model immediately through the Poolside API or via OpenRouter. For local deployment, head over to Hugging Face to download the official weights and quantized checkpoints.

If you are looking for a managed library experience, the model is also available via Baseten. We recommend starting with the FP8 checkpoints for a great balance of precision and speed on NVIDIA hardware. Check the official Poolside documentation for specific implementation recipes for vLLM and TensorRT-LLM.

• API: Poolside API & OpenRouter
• Weights: Hugging Face (poolside/Laguna-XS-2.1)
• Local Inference: Ollama, vLLM, llama.cpp (GGUF soon)

API Pricing — Input: 0.10 / Output: 0.20 / Context: 256000

Sources