Laguna S 2.1
Points and comments are a snapshot, not live.
Poolside releases Laguna S 2.1, a 118B MoE model with 8B active parameters.
Laguna S 2.1 is a Mixture-of-Experts model with 118B total parameters and 8B activated per token, supporting up to 1M token context. It went from training to launch in under nine weeks. On Terminal-Bench 2.1 it scores 70.2%, competitive with models many times its size. The model excels at long-horizon coding tasks, scoring 40.4% on DeepSWE. Poolside attributes gains to improved persistence and verification behavior rather than increased intelligence. The model is available as open weights with two thinking modes (off and max). All evaluation trajectories are published at trajectories.poolside.ai.
What commenters are saying
Commenters are impressed by the model's performance at its size, calling it a "sweet spot" for self-hosted hardware like Strix Halo and DGX Spark. Some report successful use with pi.dev and llama.cpp, with one user moving from Claude Code to Laguna S 2.1. Others caution that thinking must be enabled to see good results, and that the model can loop or overthink on complex tasks. Several commenters note the importance of the DFlash drafter for speculative decoding. The model's competitiveness with DeepSeek V4 Flash is highlighted as a welcome US counterweight to Chinese labs.