Poolside drops Laguna S 2.1, an open-weight coding model that beats rivals 10x its size

2026-07-21 · Source: VentureBeat · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Robotics & Autonomous Systems · Depth: Advanced, medium

Summary

Poolside, a San Francisco AI lab, has released Laguna S 2.1, its most capable open-weight coding model to date, available on Hugging Face under the OpenMDW-1.1 license. This 118-billion-parameter Mixture-of-Experts (MoE) system activates only 8 billion parameters per token and supports a 1 million token context window. Benchmarks show Laguna S 2.1 scoring 70.2% on Terminal-Bench 2.1, outperforming larger models like DeepSeek-V4-Pro-Max (1.6T parameters, 64.0%), Inkling (975B parameters, 63.8%), and Nemotron 3 Ultra (550B parameters, 56.4%). It also achieved 78.5% on SWE-Bench Multilingual and 59.4% on SWE-Bench Pro. Developed in under nine weeks using 4,096 Nvidia H200 GPUs, this release aims to provide a competitive Western open-weight alternative amidst a market dominated by Chinese models, emphasizing transparency and cost-efficiency for enterprise agentic coding tasks.

Key takeaway

For AI Engineers and ML Architects evaluating open-weight coding models for self-hosted agentic workloads, Poolside's Laguna S 2.1 offers a compelling Western alternative. Its sparse MoE architecture enables cost-effective inference, making enterprise-scale agentic applications more affordable. You should consider its strong benchmark performance against much larger models and the unprecedented transparency of its evaluation trajectories when selecting a foundational model for secure, on-premise deployments, especially given its broad ecosystem support and rapid iteration cycle.

Key insights

Poolside's Laguna S 2.1 demonstrates that smaller, transparent open-weight models can outperform larger rivals on agentic coding tasks.

Principles

• Radical transparency builds trust in AI benchmarks.
• Sparse MoE architectures reduce inference costs significantly.
• Model "working habits" are crucial for agentic performance.

Method

Poolside improved model performance by focusing on "working habits" like verification, persistence, and avoiding early declarations of victory during training.

In practice

• Deploy Laguna S 2.1 on a single Nvidia DGX Spark.
• Utilize 4-bit GGUF files for local model execution.
• Examine full benchmark trajectories for deeper model understanding.

Topics

• Open-weight AI
• Coding Models
• Mixture-of-Experts
• AI Benchmarking
• Enterprise AI Agents
• Model Transparency

Code references

• sgl-project/sglang
• ggml-org/llama.cpp

Best for: NLP Engineer, CTO, VP of Engineering/Data, AI Engineer, Machine Learning Engineer, AI Architect

Related on AIssential

• Poolside's Laguna S 2.1 is a small open-weight coding model that punches well above its size — Poolside's Laguna S 2.1 demonstrates that persistence and verification in agentic coding models can rival raw parameter scale.
• This Open-Weight Coding Model Makes Local Agents Real on a 128GB Mac — Poolside's Laguna S 2.1 makes powerful open-weight coding agents viable on local 128GB Mac hardware.
• Laguna S 2.1: The 118B Open AI Coding Model beats Inkling, DeepSeek — Behavioral training and MoE architecture enable compact AI coding models to outperform larger, less optimized counterparts.
• Laguna S 2.1 Open Weight Model Test | Coding with OpenCode, Frontend & CSS Design, Game Dev — Benchmarks alone do not guarantee practical coding performance for large language models.
• My Free-Model Swarm Runs 50–200× Leaner Than Me. I Reviewed the Math — and Cut the Number Down. — MoE architectures make open-weight models 10-200x more energy-efficient than dense frontier models by activating fewer parameters per token.
• [AINews] "Laguna S 2.1 Released: Cheaper than Deepseek v4 Flash, Better than V4 Pro" — Open-source models are critical for defensive AI and cost-efficiency, while agent platforms mature with advanced orchestration and evaluation tools.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by VentureBeat.