Model release notes - Poolside

Laguna S 2.1

Initial Laguna S 2.1 release.

• Context window: 1M tokens
• Model type: Mixture of Experts (MoE)
• Supported modes: Agentic, Chat
• Modality: Text-to-text
• Reasoning: Native reasoning with thinking off or max thinking

Laguna S 2.1 is a 118B total parameter Mixture-of-Experts model with 8B activated parameters per token, built for agentic coding and long-horizon work. It sits between Laguna XS 2.1 and Laguna M.1 in the Laguna model family. Use Laguna S 2.1 when you want an agent to:

• Work through long-running coding and research tasks
• Hold onto a goal across extended tool use and validation loops
• Recover from failed approaches and continue making progress
• Run an open-weight model locally on high-memory hardware, including a single NVIDIA DGX Spark

Highlights

• Long-context agentic work: Supports up to 1M tokens of context for long-horizon coding and research workflows.
• Native reasoning support: You can run requests with thinking off or with max thinking.
• Efficient sparse architecture: Has 118B total parameters and activates 8B parameters per token.
• Open-weight release: Released under OpenMDW-1.1 with BF16 weights and FP8, INT4, NVFP4, GGUF, and MLX quantizations.

Compatibility notes

• Designed for Poolside Agent CLI workflows and OpenAI-compatible API integrations.
• Download the weights from the Laguna S 2.1 Hugging Face model card.
• Run locally through pool, vLLM, SGLang, Ollama, llama.cpp, MLX, or NVIDIA TensorRT-LLM, or access through OpenRouter and the Poolside API.
• Text-to-text and does not support vision inputs.

Evaluation and limitations

• Read the Laguna S 2.1 release blog post for benchmark details, evaluation methodology, example trajectories, and known limitations.
• Full final-evaluation trajectories at trajectories.poolside.ai.
• Supports two thinking settings: off and max, with max thinking enabled by default; low, medium, and high effort controls are not available in this release.

Laguna XS 2.1

Incremental update to Laguna XS.2 with native reasoning support and stronger performance on agentic coding benchmarks.

• Context window: 256K tokens
• Model type: Mixture of Experts (MoE)
• Supported modes: Agentic, Chat
• Modality: Text-to-text

Laguna XS 2.1 is a 33B total parameter Mixture-of-Experts model with 3B activated parameters per token, designed for agentic coding and long-horizon work on local machines. It is compact enough to run on a Mac with 36 GB of RAM. Compared with Laguna XS.2, it improves on Poolside’s reference benchmarks for multilingual coding and terminal-style tasks, and adds native reasoning support with interleaved thinking between tool calls.

Improvements

• Native reasoning support: Interleaved thinking between tool calls, with the ability to enable or disable thinking per request.
• Better multilingual coding results: Improves SWE-bench Multilingual by +5.4% (57.7% to 63.1%) over XS.2.
• Better terminal task execution: Improves Terminal-Bench 2.0 by +1.8% (35.7% to 37.5%) over XS.2.
• Lower memory per token: KV cache quantized to FP8.

Compatibility notes

• Designed for Poolside Agent CLI workflows and OpenAI-compatible API integrations.
• Text-in, text-out only and does not support vision inputs.

Laguna M.1

Initial Laguna M.1 release.

• Context window: 256K tokens
• Model type: Mixture of Experts (MoE)
• Supported modes: Agentic, Chat
• Modality: Text-to-text

Laguna M.1 is a Laguna family model for agentic work. It is strongest on coding tasks that require multiple steps, tool use, and validation, such as exploring a codebase, editing files, running tests, and iterating on a fix.

Improvements

• Stronger coding performance: Improves SWE-bench Verified from 55.6% in Malibu 2.2 to 65.4%.
• Better multilingual coding results: Improves SWE-bench Multilingual from 31.1% in Malibu 2.2 to 57.4%.
• Better agentic task execution: Improves Terminal-Bench 2.0 from 16.9% in Malibu 2.2 to 32.7%.

Tips for prompting

• Give Laguna M.1 a clear task with the specific outcome you want.
• Include relevant context such as file paths, error messages, failing tests, or reference material.
• State any important constraints up front, such as coding standards, files to avoid, or commands the agent should run.

Compatibility notes

• Designed for Poolside Agent CLI workflows and OpenAI-compatible API integrations.
• Text-to-text and does not support vision inputs.
• Performance depends on the quality of the instructions and context provided, tools available to the agent, and whether the environment supports validation steps such as tests or executable checks.