Poolside has released its new Laguna S 2.1 model based on the Mixture-of-Experts (MoE) architecture, demonstrating results comparable to flagship solutions like Gemini Flash 3.6, but with significantly lower computational costs.
What Happened
The Laguna S 2.1 model possesses 118 billion parameters; however, thanks to the MoE architecture, only 8 billion parameters are activated during the generation of each token. In the Terminal-Bench 2.1 benchmark, the model scored 70.2%, placing it on par with much larger models. Additionally, Laguna S 2.1 supports a context window of up to 1 million tokens.
Context
In modern AI agent development, the focus is shifting from simply increasing the total number of parameters to improving the quality of model behavior. This includes mechanisms such as persistence, verification, and the use of test-time compute through specialized modes like "Max thinking."
Why It Matters for the Industry
The release of Laguna S 2.1 sets a new standard for efficiency in specialized AI agents for programming. It marks a transition from the race for model size to the race for optimization via MoE and the efficient use of computational resources during inference.
Why It Matters for Users
For developers and companies, this means the emergence of faster, cheaper, and more specialized AI agents capable of solving complex engineering tasks and independently fixing code errors, working effectively with massive codebases without the need for excessive computing power.
What Is Not Yet Known / Limitations
At this time, detailed data regarding inference costs, latency, security, and API availability for industrial implementation is unavailable.
Sources
Author
Look at AI, Editorial Team