Poolside Open-Sourced a 118B Coder That Solved an Erdos Problem
Poolside was supposed to be the enterprise-contract play, the lab that sells into defense and banks and never shows you the model. On July 21 they released Laguna S 2.1 with open weights on Hugging Face under OpenMDW-1.1, and the numbers are not a courtesy release.
It is 118B parameters mixture-of-experts with 8B activated per token, 1M context, built for long-horizon agentic coding. Terminal-Bench 2.1 at 70.2%, which puts it 11th overall against everything. SWE-Bench Multilingual 78.5%. SWE-Bench Pro 59.4%. DeepSWE v1.1 40.4% in thinking mode. Poolside's own framing is that it holds its own against models many times its size on long-horizon benchmarks, and with 8B active parameters that claim is checkable rather than marketing. They also say the whole thing went from pre-training start to launch in under nine weeks.
The demos are the kind that are hard to fake. It built a browser engine from scratch. It optimized complex codebases. And it independently rediscovered a proof to Erdos problem #397, which is not a coding benchmark at all and is the sort of result that either happened or did not.
Then the pricing, which is where this gets loud. Free weights on Hugging Face. Free chat at chat.poolside.ai with no login. A free OpenRouter endpoint at 256K context. And paid hosting on OpenRouter at $0.10 per million input and $0.20 per million output with the full 1M context. Ten cents. For a million tokens of context on a model that ranks eleventh on Terminal-Bench. vLLM, Ollama, Baseten and Vercel AI Gateway all have it.
The sparse-MoE trick is what makes that price possible, and it is now clearly the winning architecture for agent work specifically: you pay for 8B of compute per token and get frontier-adjacent long-horizon behavior. If your agent loop was priced out of a million-token context window a month ago, that constraint is gone. Details at poolside.ai/blog/introducing-laguna-s-2-1
← Back to all articles
It is 118B parameters mixture-of-experts with 8B activated per token, 1M context, built for long-horizon agentic coding. Terminal-Bench 2.1 at 70.2%, which puts it 11th overall against everything. SWE-Bench Multilingual 78.5%. SWE-Bench Pro 59.4%. DeepSWE v1.1 40.4% in thinking mode. Poolside's own framing is that it holds its own against models many times its size on long-horizon benchmarks, and with 8B active parameters that claim is checkable rather than marketing. They also say the whole thing went from pre-training start to launch in under nine weeks.
The demos are the kind that are hard to fake. It built a browser engine from scratch. It optimized complex codebases. And it independently rediscovered a proof to Erdos problem #397, which is not a coding benchmark at all and is the sort of result that either happened or did not.
Then the pricing, which is where this gets loud. Free weights on Hugging Face. Free chat at chat.poolside.ai with no login. A free OpenRouter endpoint at 256K context. And paid hosting on OpenRouter at $0.10 per million input and $0.20 per million output with the full 1M context. Ten cents. For a million tokens of context on a model that ranks eleventh on Terminal-Bench. vLLM, Ollama, Baseten and Vercel AI Gateway all have it.
The sparse-MoE trick is what makes that price possible, and it is now clearly the winning architecture for agent work specifically: you pay for 8B of compute per token and get frontier-adjacent long-horizon behavior. If your agent loop was priced out of a million-token context window a month ago, that constraint is gone. Details at poolside.ai/blog/introducing-laguna-s-2-1
Comments