Now loading...
Laguna S 2.1, the latest advancement in AI development, has been unveiled as a new model designed for improved long-term reasoning and performance. This model boasts a total of 118 billion parameters and incorporates a mixture of experts (MoE) design that activates 8 billion parameters per token. Notably, it supports a context window of up to 1 million tokens, enabling it to handle complex tasks efficiently. The development timeline for Laguna S 2.1 was rapid, taking less than nine weeks from the start of training to launch. It has performed comparably in coding benchmarks against larger models, further emphasizing its capabilities.
The model demonstrates impressive benchmark performance, achieving a score of 70.2% on Terminal-Bench 2.1 in the agent harness with its reasoning features activated. This performance showcases its potential for intricate work on personal computing devices. Full evaluation trajectories for all benchmarks will be made publicly available at trajectories.poolside.ai.
Laguna S 2.1 has been recognized for punching above its weight class in the field of agentic coding models, even outperforming larger competitors in specific tasks. Its strong results stem from a significant amount of verification and persistent problem-solving behaviors, which have been refined through its training and evaluation methodologies. Recent tests indicate that the model is capable of sustained, coherent thinking over extended periods, marking a notable improvement over its predecessors.
One specific case study highlighted how Laguna S 2.1 could autonomously create a browser engine from scratch. Without any visual capabilities, it successfully built a working HTML/CSS engine, demonstrating its resourceful problem-solving approach by validating its work against established standards.
Another demonstration of its capabilities involved optimizing its own evaluation harness, where it improved processing speed by 5.2% while drastically reducing memory allocation. This case reflects the model’s potential for meaningful engineering and research applications.
In the sphere of mathematical problem-solving, Laguna S 2.1 excelled by independently rediscovering a proof for Erdős problem #397. Its mathematical prowess indicates that it holds significant promise in various cognitive tasks beyond just coding.
Laguna S 2.1 offers two thinking modes for users: a basic mode and a maximum thought mode, which enhances its performance on relevant benchmarks. The current rollout ensures that users can access the model quickly and start using its powerful features immediately. As a result of its capabilities and rapid deployment, Laguna S 2.1 is positioned to contribute significantly to advancements in AI and software engineering.
This innovative model is now accessible via platforms such as Hugging Face and other hosting services, ensuring that developers can deploy it across multiple environments effectively.
