Laguna XS 2.1

Laguna XS 2.1 is a 33B total parameter Mixture-of-Experts (MoE) coding agent model with 3B activated parameters per token, specifically engineered for high-performance agentic coding and long-horizon tasks. It features an efficient architecture with mixed sliding-window and global attention, native interleaved reasoning, and support for a 256K context window, making it highly capable for local execution and complex software engineering workflows.

Laguna XS 2.1 is a 33B total parameter Mixture-of-Experts (MoE) coding agent model with 3B activated parameters per token, specifically engineered for high-performance agentic coding and long-horizon tasks. It features an efficient architecture with mixed sliding-window and global attention, native interleaved reasoning, and support for a 256K context window, making it highly capable for local execution and complex software engineering workflows.

• 0133B total parameter MoE model with 3B activated parameters
• 02256K-token context window with native interleaved reasoning support
• 03Mixed sliding-window (512) and global attention (3:1 ratio)
• 04FP8 KV cache quantization for reduced memory usage
• 05Optimized for local deployment via Ollama, vLLM, and llama.cpp
• 06Supports speculative decoding with the DFlash draft model