Poolside's Laguna S 2.1 ships FP8 with a natively trained 1M-token window
The updated checkpoint mixes sliding-window and global attention at a 3:1 ratio across 48 layers, quantises the KV cache to FP8, and supports interleaved reasoning between tool calls. The million-token context is natively trained rather than extended after the fact.
The architectural detail worth pulling out is the 3:1 ratio. Three sliding-window layers to one global layer is a direct trade of long-range attention for memory bandwidth, and it is the mechanism that makes a million-token window affordable rather than merely advertised.
Natively trained at 1M is the other claim that changes what the number means. Context windows extended by post-hoc interpolation tend to degrade well before the stated limit; a model trained at the length it advertises does not carry that particular asterisk, though it says nothing about whether attention is usefully distributed across the span.
Interleaved thinking between tool calls is the agentic tell. It signals the checkpoint is aimed at long-horizon harness use rather than single-turn completion, which is where the whole coding-model field has converged. Published scores reference Terminal-Bench 2.1, DeepSWE via Artificial Analysis and SWE Atlas.
Hugging Face — poolside/Laguna-S-2.1-FP8 model card → · Hugging Face — poolside/Laguna-S-2.1-NVFP4 model card →