Key takeaways

• Laguna S 2.1 is Poolside's 118B total parameter Mixture-of-Experts (MoE) coding model, with 8B activated parameters per token, designed for agentic coding and long-horizon software engineering tasks.
• It features a sparse MoE architecture with 256 routed experts, mixed sliding-window and global attention, and supports a context window of up to 1 million tokens.
• The model demonstrates strong mathematical problem-solving capabilities, achieving 96.67% accuracy on the MATH-500 dataset, particularly excelling in algebra and number theory, and showing robust problem-solving in fundamental physics and chemistry.
• A key limitation is its sensitivity to output formatting rules, which frequently led to score reductions despite mathematically correct answers in the MATH-500 evaluation, and struggles with nuanced interpretations in molecular biology, genetics, and complex organic chemistry reactions.
• While capable of complex calculations and step-by-step reasoning, it occasionally falters on nuanced interpretations or specific details, particularly in chemistry and biology, and improvements are needed in handling multi-step organic reactions and complex biological experimental design questions.

Technical specifications

Context length

1M tokens

Max output

—

Architecture

MoE

Modalities

text->text

Parameters

~118B

Open weights

Yes

Licence

Open

Regions

USA

Vendor

Poolside

Provider model key

poolside/laguna-s-2.1