Poolside Laguna S 2.1

Quick Answer

Laguna S 2.1 is Poolside’s open-weight coding model, released July 21, 2026. It’s a 118-billion-parameter Mixture-of-Experts system that activates roughly 8 billion parameters per token, supports up to a 1M-token context window, and is built specifically for agentic, multi-step software engineering. Poolside ships it under its own permissive license, OpenMDW-1.1, and the weights are free to download.

It sits alongside a smaller companion model, Laguna XS 2.1 (33B total parameters, 3B active), released July 2, 2026, also free on Hugging Face and OpenRouter.

Laguna S 2.1 is the model. pool, Poolside’s terminal coding agent, is a separate tool that can run on Laguna models or other providers.

What Laguna S 2.1 Is Best For

• Agentic coding on modest hardware: multi-step software-engineering tasks on a single GPU
• Long-context coding work: up to 1M tokens for full-repository context
• Open-weight deployment: self-hosting without needing a large multi-GPU cluster
• Cost-conscious coding agents: a smaller, cheaper alternative to running a frontier-scale model for coding tasks

Key Facts (Verified)

• Developer: Poolside AI
• Released: July 21, 2026
• Type: Open-weight Mixture-of-Experts model, 118B total parameters, ~8B active per token
• Context window: up to 1M tokens
• License: OpenMDW-1.1, a permissive open-weight license introduced by Poolside
• Availability: Hugging Face, OpenRouter
• Companion model: Laguna XS 2.1 (33B total / 3B active), released July 2, 2026

Poolside’s published benchmarks on Terminal-Bench 2.1 and SWE-Bench Pro show Laguna S 2.1 matching or beating several larger models, including DeepSeek-V4-Flash, Nvidia’s Nemotron 3 Ultra, and Thinking Machines’ Inkling. These are vendor-reported results, so confirm on your own coding workflows before relying on the comparison.

How to Access It

Self-hosted: download the weights from Hugging Face and run on a single capable GPU, no multi-GPU cluster required.

Hosted: use it through OpenRouter on a pay-per-use basis, no infrastructure setup needed.

As a coding agent: pair it with pool or another ACP-compatible coding agent for terminal or editor-based workflows.

Honest Limitations

• Still a serious model to self-host: “single GPU” means a capable one, not a laptop
• New license: OpenMDW-1.1 is Poolside’s own license, read the terms before commercial use
• Vendor benchmarks: Terminal-Bench and SWE-Bench Pro results come from Poolside, verify independently
• Narrower focus than general-purpose frontier models: built for coding, not a general chat replacement

Alternatives Worth Knowing

• GLM 5.2, a larger open-weight model also strong at coding, needs more hardware
• Kimi K3, Moonshot AI’s much larger open-weight frontier model
• DeepSeek, another open-leaning coding-capable model
• Tinker, Thinking Machines’ fine-tuning platform, useful if you want to customize an open model like this one
• pool, Poolside’s own coding-agent CLI

For the bigger picture on open-weight models, see Frontier Open Models Explained and Open Models vs Closed Models.