Laguna

Two sparse MoE coding models from Poolside — Laguna XS 2.1 and Laguna S 2.1.

Laguna is the coding-focused MoE family in the matrix. XS is a 33B total / 3B active model sized for a single local machine; S is a 118B total / 8B active model in the same recipe. Both use sliding-window and global attention together with a large expert pool, and both compile through the same HIP EP pipeline as the rest of the matrix.

Laguna XS 2.1

  • Kind: Language model
  • Task: Text generation, coding
  • Parameters: 33B total, 3B active
  • Architecture: Sparse mixture of experts
  • Quantization: int4 RTN, group 128
  • Base model: poolside/Laguna-XS-2.1
  • Validation status: New in v0.5.3
  • Measured in v0.5.3: 66.38 tokens/s, 0.89 s to first token, at a 2048-token prompt — all three lengths

Laguna S 2.1

Added in v0.5.3. A 118B MoE coding model that still fits the same Strix Halo validation path as the rest of the matrix.

  • Kind: Language model
  • Task: Text generation, coding
  • Parameters: 118B total, 8B active
  • Architecture: Sparse mixture of experts
  • Quantization: int4 RTN, group 128
  • Base model: poolside/Laguna-S-2.1
  • Validation status: New in v0.5.3
  • Measured in v0.5.3: 33.61 tokens/s, 2.59 s to first token, at a 2048-token prompt — all three lengths