Laguna S 2.1: The 118B MoE model outpacing giants on coding benchmarks
A 118-billion-parameter model just outperformed systems five to twenty times larger on long-horizon coding tasks—and it does so while fitting on a single NVIDIA DGX Spark. Poolside’s Laguna S 2.1, a sparse Mixture-of-Experts model that activates only 8 billion parameters per token, delivers 78.5% on SWE-Bench Multilingual and 70.2% on Terminal-Bench 2.1 with its default “max” thinking mode enabled. The weights are openly available under the OpenMDW-1.1 license on Hugging Face, with official conversions for GGUF and MLX to broaden accessibility.
How it stays lightweight yet powerful
Unlike dense models that scale compute with every parameter, Laguna S 2.1 routes only about 6.8% of its total 118 billion weights per token through the network. All parameters remain resident in memory, but the sparse activation keeps inference fast and cost-efficient. The team trained it on 4,096 NVIDIA H200 GPUs in under nine weeks and is the first Poolside release to use reinforcement learning in FP8 precision. Official weights are published in BF16, FP8, INT4, and NVFP4, plus DFlash draft models for even faster inference.
Thinking mode matters—and costs tokens
Laguna S 2.1 ships with two modes: off and max. The default max mode lets the model self-adjust its compute budget during reasoning, lifting Terminal-Bench 2.1 from 60.4% to 70.2% and DeepSWE v1.1 from 16.5% to 40.4%. Those gains come at the expense of longer trajectories—DeepSWE runs averaged 249k completion tokens in thinking mode versus 99k without. Poolside also published three unedited trajectories showing the model building a working HTML/CSS browser engine from scratch in a 50-minute session with 181 steps and no human intervention.
Why it matters
Laguna S 2.1 demonstrates that agentic coding performance is no longer the sole domain of trillion-parameter closed systems. Its open weights and modest hardware footprint lower the barrier for researchers and startups to build advanced coding agents. Yet the gap on some benchmarks—especially DeepSWE v1.1—suggests room for further refinement. For practitioners, the model offers a practical path to high-quality code generation without the prohibitive cost of frontier systems.
Source: MarkTechPost. AI-assisted editorial synthesis — TechnoExpress.