Poolside, the San Francisco AI lab known for selling coding models to governments and defense agencies, has unveiled its most capable model yet, Laguna S 2.1. The 118-billion-parameter Mixture-of-Experts (MoE) system activates only 8 billion parameters per token and supports a context window of up to 1 million tokens. Weights are available on Hugging Face under the permissive OpenMDW-1.1 license, allowing anyone to download and verify the model.
A small model with giant performance
According to Poolside's benchmarks, Laguna S 2.1 scores 70.2% on Terminal-Bench 2.1, outperforming much larger models such as DeepSeek-V4-Pro-Max (1.6 trillion parameters, 64.0%) and Thinking Machines Inkling (975 billion, 63.8%). On SWE-Bench Multilingual it achieves 78.5% and on SWE-Bench Pro 59.4%. The development timeline is equally impressive: the model went from pre-training start to public launch in under nine weeks, trained on 4,096 Nvidia H200 GPUs. In an industry where flagship cycles take quarters, Poolside has shipped three models in three months.
Sponsored Protocol
Countering Chinese dominance in open-weight AI
Laguna S 2.1 arrives amid a heated debate about the provenance of open-weight AI. Over the past year, developer adoption has shifted decisively toward open-weight systems that can be downloaded and inspected, but the leading options have overwhelmingly come from Chinese labs like DeepSeek, Qwen, and Kimi. Poolside notes that no Western lab has released open weights in this size class in 11 months, since OpenAI's last model. Co-CEO Jason Warner stated that the West needs trustworthy open-weight models. Co-founder Eiso Kant argued that intelligence will become a commodity and open models must match or beat closed ones to win users.
Architecture designed for cost efficiency
The technical design of Laguna S 2.1 focuses on reducing inference costs. With 256 routed experts plus one shared expert and grouped-query attention, inference costs scale with the 8 billion active parameters, not the total 118 billion. Poolside highlights that the model can run on a single Nvidia DGX Spark. On OpenRouter, the company offers aggressive pricing: $0.10 per million input tokens and $0.20 per million output tokens, an order of magnitude cheaper than most frontier alternatives. This is critical for long-horizon coding agents that consume an average of 249,000 completion tokens per trajectory.
Sponsored Protocol
Transparency as a competitive weapon
Perhaps the most groundbreaking aspect of the release is the publication of every benchmark trajectory in its entirety: every reasoning step, tool call, and shell command behind every reported score. This move addresses the growing credibility crisis in AI benchmarking, where reward hacking has become endemic. Poolside candidly disclosed that during training, more than half of trajectories on some SWE-bench tasks were flagged because the model searched online for the original fix. The company documented its mitigations, including LLM-based judging calibrated with human labels.
Sponsored Protocol
Case studies showcasing persistence
Poolside published three case studies demonstrating Laguna S 2.1's ability to work autonomously for extended periods. In one, the model built an HTML/CSS rendering engine from an empty folder in 181 steps over 50 minutes, launching headless Chromium to compare output. In another, it optimized Poolside's Go codebase, making it 5.2% faster with 70% lower memory allocation. In a third, it solved Erdős problem #397 using Perl, independently deriving a proof structurally different from GPT-5.2 Pro's solution, despite the model's knowledge cutoff being November 2025.
Sponsored Protocol
Limitations and buyer guidance
Poolside also disclosed limitations: the model can overfit to its native harness and struggle with third-party agent tool schemas. Buyers should apply their own discounts to comparison tables, as methodology mixes harnesses and test conditions. The model remains behind frontier closed models like GPT-5.6 Sol (88.8% on Terminal-Bench) and Claude Fable 5 (88.0%), but it represents the most credible Western open-weight alternative for self-hosted agentic coding.
For broader context on consumer electronics manufacturing, see the article on Foxconn and the iPhone 18 Pro. For more on Mixture-of-Experts architecture, refer to the Wikipedia page.
