Poolside releases Laguna S 2.1, 118B-parameter open-weight model matching closed competitors

Poolside released Laguna S 2.1, a 118-billion-parameter open-weight Mixture-of-Experts model for agentic coding, on July 21, 2026. The model activates only 8 billion parameters per token, supports a 1-million-token context window, and is available on Hugging Face under the OpenMDW-1.1 license. Weights are available in multiple formats (BF16, FP8, INT4, NVFP4, GGUF, MLX) and the model runs on a single NVIDIA DGX Spark. Poolside trained the model in under nine weeks on 4,096 H200 GPUs.

On Terminal-Bench 2.1 (agentic terminal tasks), Laguna S 2.1 scored 70.2% with thinking mode enabled—beating DeepSeek V4 Pro Max (1.6T total, 49B active) at 64.0% and NVIDIA Nemotron 3 Ultra (550B, 55B active) at 56.4%. On SWE-Bench Pro, it achieved 59.4% versus DeepSeek V4 Pro Max at 55.4%. Most dramatically, on DeepSWE (harder multi-file tasks), Laguna S 2.1 scored 40.4% while DeepSeek V4 Pro Max scored just 9.0%—a 31.4-point gap that suggests Poolside's emphasis on verification and persistence behaviors outperforms raw model size.

Laguna S 2.1 is the first Western open-weight model in its size class released in 11 months—the last comparable release was OpenAI's gpt-oss-120b in August 2025. Poolside co-CEO Jason Warner framed it as a response to Chinese dominance in open-weight systems (DeepSeek, Qwen, Kimi) and marketed it as "the West needs open-weight models it can trust." Pricing is $0.10/$0.20 per million input/output tokens on OpenRouter, or free locally if you have hardware.

For developers: Laguna S 2.1 demonstrates that parameter-efficient training (8B active, trained in <9 weeks) can outcompete larger dense and MoE models on real agentic benchmarks, validating the efficiency thesis. However, Poolside acknowledges the model is "not yet at the frontier"—closed-source leaders (OpenAI, Anthropic) still lead by ~10–15 points on Terminal-Bench. The test now is whether Terminal-Bench scores translate to production performance on messy, non-curated codebases, and whether Poolside can maintain the pace against Chinese labs that are also improving rapidly.