🤖 Poolside introduces Laguna S 2.1: A powerful model for coding
Poolside has released Laguna S 2.1, featuring a Mixture-of-Experts (MoE) architecture with 118 billion parameters. By activating only 8 billion parameters per token, the model operates extremely efficiently with a context window of up to 1 million tokens. In the Terminal-Bench 2.1 benchmark, it scored 70.2%, which is comparable to Gemini Flash 3.6.
🌍 Laguna S 2.1 sets a new standard for efficiency in agentic coding, shifting the focus from simple parameter scaling to behavioral quality (persistence, verification) and the use of test-time compute in 'Max thinking' mode.
👤 This represents a path toward the emergence of fast and inexpensive specialized AI agents for programming, capable of solving complex engineering tasks and self-correcting errors without massive computational costs.