Poolside's Laguna S 2.1 crushes DeepSeek on DeepSWE with one-sixth the active parameters

Poolside dropped Laguna S 2.1 on Thursday, a 118-billion-parameter open-weight coding model that punches wildly above its weight class. It’s a Mixture-of-Experts architecture that only activates about 8 billion parameters per token—roughly 6.8% of the total. That sparsity lets the model behave like a much larger system while keeping inference costs manageable. The entire training run, from start to launch, took under nine weeks on 4,096 NVIDIA H200 GPUs.

The real headline is DeepSWE v1.1. Laguna S 2.1 scores 40.4% on that benchmark with thinking enabled, while DeepSeek-V4-Pro-Max manages just 9.0%. That’s a staggering gap, and Laguna does it with roughly one-sixth the active parameters. On Terminal-Bench 2.1, the model hits 70.2%, and it leads the pack on SWE-Bench Multilingual at 78.5%. Closed frontier systems like Claude Fable 5 still hold the top spots overall, but Poolside’s claim is about dominating the open-weight, disclosed-size category. They’re not wrong.

The model ships with two thinking modes: off and max. Max is the default, and it’s where the magic happens—at a cost. DeepSWE trajectories in max mode average 249,000 completion tokens compared to 99,000 without. That’s a lot of compute, but Poolside published unedited trajectories showing the model running autonomously for nearly an hour, building a working browser engine from scratch and validating its output against headless Chromium. No human intervention, 181 steps, 50 minutes. In another run, it optimized Poolside’s own agent harness, delivering a 5.2% speedup and slashing memory allocation by 71%.

Deployment is surprisingly straightforward for a model this size. At 4-bit precision, the weights fit in about 59 GB—comfortable on a single NVIDIA DGX Spark with 128 GB of unified memory. FP8 pushes it to roughly 118 GB, still within a single Spark or H200. BF16 requires two linked Sparks. Poolside worked directly with NVIDIA on inference optimization and shipped day-one support for vLLM, SGLang, and Ollama. Hosted access is available through OpenRouter, free at 256K context and paid at the full 1M-token window. It’s their third model in under three months. The pace is relentless, and the trajectory is clear: agentic coding is the new battleground, and open-weight models are now legitimate contenders.