Laguna-XS-2.1

### TL;DR

Laguna-XS-2.1 is a 33B parameter Mixture-of-Experts (MoE) model designed for agentic coding and complex, long-horizon tasks. Featuring 3B activated parameters per token, it provides enhanced performance on benchmarks like SWE-bench Multilingual while remaining efficient enough to run on local hardware.

Key Insights & Metrics

Key Features

• 33B total parameters with 3B activated per token
• Mixed sliding window and global attention layout (3:1 ratio)
• Native reasoning support with interleaved thinking capabilities
• FP8 KV cache quantization for memory efficiency
• Agentic coding optimization with strong terminal-task performance

Discussion