Poolside Unveils Laguna S 2.1: a Game-changing Coding Model
WHO:
VentureBeat - Poolside, a San Francisco AI lab, co-CEOs Jason Warner and Eiso Kant.
Techmeme - Iain Martin, reporting for Forbes.
WHAT:
VentureBeat - Laguna S 2.1 is reported to match or beat open models several times its size on agentic coding tasks.
VentureBeat - Laguna S 2.1 is claimed to outperform DeepSeek-V4-Pro-Max, a 1.6-trillion-parameter model that scored 64.0, and Thinking Machines' 975-billion-parameter Inkling, which scored 63.8.
VentureBeat - The model's design allows it to run on a single Nvidia DGX Spark.
VentureBeat - Poolside published the complete, unedited trajectory of every trial in its final benchmark runs, addressing credibility issues in AI benchmarking.
VentureBeat - The model's construction is structurally different from the earlier published solution to Erdős problem #397, which was first solved by GPT-5.2 Pro in January 2026.
VentureBeat - Poolside released its coding model, Laguna S 2.1, which has 118 billion parameters and activates only 8 billion parameters per token.
VentureBeat - Laguna S 2.1 supports a context window of up to 1 million tokens and scores 70.2% on Terminal-Bench 2.1.
VentureBeat - The model was trained on 4,096 Nvidia H200 GPUs and went from pre-training to public launch in under nine weeks.
VentureBeat - Laguna S 2.1 is available on Hugging Face under the OpenMDW-1.1 license.
VentureBeat - The model scored 78.5% on SWE-Bench Multilingual and 59.4% on SWE-Bench Pro's public dataset.
VentureBeat - Poolside published the complete, unedited trajectory of every trial in its final benchmark runs.
WHERE:
VentureBeat - San Francisco, California, where Poolside is based.
WHEN:
VentureBeat - The model began pre-training on May 22 and was publicly launched in under nine weeks.
VentureBeat - The next, larger Laguna model began pre-training last week.
WHY:
VentureBeat - Poolside aims to provide open-weight models that can be trusted, run, and built upon by Western labs, responding to the dominance of Chinese open-source models.
VentureBeat - The release is part of Poolside's strategy to compete in the AI market by focusing on cost per token, self-hosting, and iteration speed.
HOW:
VentureBeat - Poolside's model is designed to consume a mean of roughly 249,000 completion tokens per trajectory on its hardest benchmark when thinking mode is enabled.
VentureBeat - Laguna S 2.1 uses a sparse Mixture-of-Experts architecture with 256 routed experts and one shared expert.
VentureBeat - The model's design allows inference costs to scale with the 8 billion active parameters rather than the total 118 billion.
VentureBeat - Poolside offers a free 256K-context endpoint and a dedicated 1M-context deployment priced at $0.10 per million input tokens and $0.20 per million output tokens.
(UPDATED July 22, 2026 01:26 UTC)
Laguna S 2.1 has been reported to match or even surpass open models several times its size on agentic coding tasks, outperforming both DeepSeek-V4-Pro-Max and Thinking Machines' Inkling in recent benchmarks. This new model is designed to run on a single Nvidia DGX Spark and consumes an average of about 249,000 completion tokens per trajectory during its most challenging tests. Additionally, Poolside has published the complete, unedited trajectory of every trial in its final benchmark runs, addressing previous credibility concerns in AI benchmarking.
--
(UPDATED July 21, 2026 23:18 UTC)
Poolside, an AI lab based in San Francisco, has just released its new coding model, Laguna S 2.1, which boasts an impressive 118 billion parameters but activates only 8 billion per token. This model supports a context window of up to 1 million tokens and achieved notable scores of 70.2% on Terminal-Bench 2.1 and 78.5% on SWE-Bench Multilingual, all while being trained on 4,096 Nvidia H200 GPUs in under nine weeks. Available on Hugging Face under the OpenMDW-1.1 license, Poolside aims to provide trusted open-weight models to compete in the AI market, focusing on cost efficiency and rapid iteration.