Poolside releases Laguna S 2.1, a 118B-parameter open-weight coding model

Poolside released Laguna S 2.1 on July 21, 2026, an open-weight foundation model built for agentic coding and long-horizon software engineering tasks, which the company is positioning as the West's most capable open-weight model to date.

What's new

Laguna S 2.1 is a Mixture-of-Experts model with 118B total parameters and roughly 8B activated per token, using 48 layers split in a 1:3 global-to-sliding-window-attention ratio (12 global attention layers, 36 sliding-window layers). The model supports a context window up to 1,048,576 tokens and offers interleaved thinking between tool calls with per-request control, a feature aimed at long, multi-step agentic coding sessions.

On benchmarks, Poolside reports Laguna S 2.1 scoring 78.5% on SWE-Bench Multilingual, 59.4% on SWE-Bench Pro (Public Dataset), and 70.2% on Terminal-Bench 2.1 — putting it competitive with models several times its size, including DeepSeek-V4-Pro-Max and NVIDIA's Nemotron 3 Ultra, according to the model's own documentation. Poolside's model card describes it as "118B total parameter Mixture-of-Experts model with 8B activated parameters per token, designed for agentic coding and long-horizon work."

The model is released under Poolside's OpenMDW-1.1 license, which permits users to "use and modify the model and associated materials freely for commercial and non-commercial purposes." Weights are available on Hugging Face in multiple formats — FP8, NVFP4, INT4, and GGUF — alongside a trained DFlash draft model for speculative decoding, and the company says quantized variants are small enough to run on a single NVIDIA DGX Spark.

Laguna S 2.1 slots into Poolside's existing model family between the smaller Laguna XS 2.1 (33B-A3B) and the larger Laguna M.1 (225B-A23B), sharing the same "laguna" architecture and the same engine integrations — vLLM, SGLang, Transformers, TRT-LLM, and llama.cpp.

Context

Poolside has built its reputation specifically around code-generation and agentic-coding models, competing in a lane increasingly crowded by both open-weight releases from Chinese labs (DeepSeek, Qwen) and closed frontier labs' coding-focused offerings. Releasing Laguna S 2.1 as open weights, rather than API-only, puts Poolside in more direct competition with DeepSeek and Qwen's open releases specifically, rather than only with closed coding assistants.

The long-context (1M token) and interleaved-thinking design choices reflect a broader industry shift toward models built explicitly for "long-horizon" autonomous coding work — multi-hour or multi-day agentic sessions rather than single-turn code completion — a category several labs have been racing to establish benchmarks and safety practices around this year.

Why it matters

A 118B-parameter open-weight model beating or matching much larger closed and open competitors on SWE-Bench Multilingual is a notable efficiency result, and the fact that quantized versions can run on a single DGX Spark lowers the bar for organizations wanting to self-host a capable coding agent rather than depend on an API. Combined with the permissive OpenMDW-1.1 license, Laguna S 2.1 gives enterprises and independent developers a credible, inspectable alternative to closed agentic-coding APIs at a moment when many are wary of sending proprietary codebases to third-party model providers.