SYLIN

Laguna S

by Poolside · Model · Language & code models

Poolside's 118B MoE coding model, 8B active, 1M context, thinking modes; 33B XS variant also released with open weights

AccessOpen weightsNeeds~235 GBDownload235 GBSize118B parametersLicenseOpenMDW-1.1

Evidence

  1. Jul 26
    Via Hermes Agent API, found and fixed an unstated seaport-app berth-draft bug, but overthought and hallucinated two extra bugs
  2. Jul 26
    Self-reported DeepSWE 40%, slightly above GLM 5.2 which is nearly seven times larger
  3. Jul 26
    In the reviewer's tests the XS variant is well behind Qwen 3.6 35B and 27B, the best medium local models
  4. Jul 22
    Laguna S 2.1: 118B MoE, 8B active per token, 1M context, native interleaved reasoning, built for long-horizon agentic coding
  5. Jul 22
    Model card chart shows Terminal-Bench 70.2 vs 64 for DeepSeek V4 Pro Max; reviewer says he doesn't trust the benchmark

Cost to run