Laguna XS.2

#10

poolside · seit 2026-04-28 · 16× · zuletzt 30. Juni 2026

25

Momentum

Laguna XS.2 is a Mixture-of-Experts language model by poolside with 33 billion total and 3 billion active parameters per token, built entirely from scratch for agentic coding and long-horizon tasks on local hardware. It is poolside's first open-weight release, trained through pre-training, post-training, and reinforcement learning stages. The architecture combines sliding-window attention in 30 out of 40 layers with 10 global attention layers (3:1 ratio) and an FP8-quantized KV cache. The model runs on a single consumer GPU or a Mac with 36 GB of RAM and natively supports interleaved reasoning between tool calls.

Preuves (16)