Laguna S 2.1 Released: A High-Performance 118B-A8B Local LLM
Laguna S 2.1, a new 118B-A8B large language model optimized for local setups, has been released. It reportedly outperforms Deepseek V4 Pro on SWE-bench and is cheaper than Deepseek v4 Flash. This release offers a highly capable, cost-effective alternative for local LLM deployments and software engineering agents. It demonstrates that sparse Mixture of Experts (MoE) models can deliver state-of-the-art coding capabilities on consumer-accessible hardware. The model features a 118B total parameter size with 8B active parameters (118B-A8B), making it suitable for setups with 64GB+ of RAM and VRAM. It achieved 78.5% on SWE-bench, 70.2% on Terminal-Bench 2.1, and 40.4% on SWE-Bench Pro.
## BACKGROUND
SWE-bench is a benchmark that evaluates AI models on their ability to resolve real-world software engineering issues from GitHub. Terminal-Bench evaluates AI agents on command-line tasks, while Mixture of Experts (MoE) architectures allow large models to run efficiently by only activating a fraction of their total parameters (active parameters) per token.