On July 21, 2026, Poolside released Laguna S 2.1, an open-weight coding model. It is a mixture-of-experts design with 118 billion total parameters and 8 billion activated per token, and it carries a 1 million token context window in both its thinking and no-thinking modes.
Poolside published pass@1 scores of 70.2 percent on Terminal-Bench 2.1, 78.5 percent on SWE-Bench Multilingual, 59.4 percent on the public SWE-Bench Pro dataset, 40.4 percent on DeepSWE v1.1, 46.2 percent on SWE Atlas codebase question answering, and 49.7 percent on Toolathlon Verified. The weights are released under the OpenMDW-1.1 license and hosted on Hugging Face, with local inference through Ollama and atomic.chat and hosted access via Baseten’s Model Library, OpenRouter, and the Vercel AI Gateway. Poolside also lists integrations with agent frameworks including Kilo, Hermes Agent, Pi, OpenCode, OpenClaw, Cline, and its own pool client.
The size figures are the point. An 8 billion active parameter model that posts these coding scores is cheap enough to serve at scale and small enough to run on a single workstation-class accelerator, which is a different deployment story from the trillion-parameter open-weight models released the same month.
For businesses, the practical consequence is that a competitive agentic coding model can now sit inside the corporate network under a permissive license, with no per-token bill and no code leaving the building. That is the argument open-weight vendors have been making for two years; the gap in capability that made it a hard sell keeps narrowing.