# Poolside releases 118B Laguna S 2.1 for local coding agents > Source: > Published: 2026-07-21 18:43:21+00:00 [Poolside](https://poolside.ai/) co-founders [Eiso Kant (@eisokant)](https://x.com/eisokant) and [Jason Warner (@jasoncwarner)](https://x.com/jasoncwarner) released Laguna S 2.1 on July 21st, an open-weight coding model built to bring long-running software agents onto hardware controlled by their users. The [118-billion-parameter Mixture-of-Experts model](https://x.com/poolsideai/status/2079613777343848465) activates 8 billion parameters for each token and supports context windows of up to 1 million tokens. Poolside released separate thinking and no-thinking modes, along with BF16, FP8, INT4 and NVFP4 weights. Official GGUF, MLX and DFlash conversions are also planned for local inference setups. The release extends a founder thesis that predates Poolside. Kant previously built source{d}, which applied machine learning to source code, and later founded engineering analytics platform Athenian. Warner ran technology at GitHub, where he worked as CTO during the development of GitHub Copilot, before spending two years as a managing director at Redpoint Ventures. The pair founded Poolside in 2023 around the view that code execution gives AI models a direct feedback mechanism: software either runs or it does not. Laguna S 2.1 is the third major size point in Poolside's current model family. The earlier Laguna XS 2.1 has 33 billion total parameters and activates 3 billion, while Laguna M.1 has 225.8 billion parameters and activates 23.4 billion. Poolside positioned the new model between them, with enough total capacity for longer agentic tasks and a lower active parameter count intended to limit inference costs. ### A local model with a benchmark caveat Poolside says Laguna S 2.1 can run on a single Nvidia DGX Spark. That claim is central to the release: a model that can be downloaded and operated on controlled infrastructure gives enterprises an alternative to sending proprietary code through a metered cloud API. The active parameter count does not eliminate the storage and memory demands of a 118-billion-parameter checkpoint. Mixture-of-Experts routing reduces the amount of the network used for each token, while the full set of weights still has to be stored or distributed across the inference system. The lower-precision INT4 and NVFP4 releases are therefore important parts of Poolside's local deployment pitch, rather than secondary conversions. Poolside reports that Laguna S 2.1 scored 70.2% on [Terminal-Bench 2.1](https://www.tbench.ai/leaderboard/terminal-bench/2.1?agents=Terminus+2), 78.5% on SWE-bench Multilingual, 59.4% on SWE-Bench Pro and 40.4% on [DeepSWE v1.1](https://deepswe.datacurve.ai/). Those results were run with [pool, Poolside's own coding-agent harness](https://github.com/poolsideai/pool), rather than a single standardized agent shared across all comparison models. The distinction matters because coding benchmarks measure an agent-model pair. Tool selection, retry policies, context management and step limits can materially change the score produced by the same underlying model. Poolside's 70.2% Terminal-Bench result was not listed on the benchmark's independently verified leaderboard when Laguna S 2.1 was released. Poolside also said its comparison table used the highest publicly referenced result for each rival model and that full methodology and sampling details would accompany the release. That leaves the performance ranking as a Poolside-reported result. The architecture and weight release are verifiable; claims that Laguna S 2.1 leads its weight class require comparable third-party runs. ### Poolside is selling control The weights are [available through Hugging Face](https://huggingface.co/poolside/Laguna-S-2.1) under the OpenMDW-1.1 license. Poolside is also distributing access through OpenRouter, Vercel's AI Gateway, Baseten and its own API, while supporting the model inside the pool terminal agent. Open distribution feeds Poolside's enterprise strategy. In May, Poolside introduced a platform for running models and agents inside customer-controlled infrastructure, with audit logs, network policies and deployment options for private clouds, on-premises systems and restricted environments. The model release gives engineers a way to test Poolside's work directly; the commercial platform addresses the governance and operational work required to deploy it across an organization. Warner described the target market in Poolside's release materials: "The West needs open-weight models it can trust, run, and build on." Poolside specifically markets its deployment stack to government, defense and regulated enterprises where source code and internal data cannot leave a controlled security boundary. Poolside says Laguna S 2.1 took less than four weeks to train on 4,000 Nvidia H200 GPUs. Poolside attributes that cadence to its Model Factory, an internal system for automating data preparation, training, evaluation and architecture experiments. A technical report published in May said Poolside built the earlier Laguna XS.2 from the start of training to release in five weeks. For Laguna S 2.1, Poolside said it reused the same pretraining data as Laguna XS 2.1. The scale increase, training-code fixes and smaller recipe changes accounted for part of the improvement, while Poolside attributed most of the capability gain to supervised fine-tuning and reinforcement learning after pretraining. Poolside also disclosed limits that matter for agent deployments. Laguna S 2.1 can spend too much time reasoning on difficult mathematics, and its first release does not provide an intermediate reasoning-effort setting between no-thinking and maximum thinking. Poolside is already training the next iteration to address those behaviors. Laguna S 2.1 gives Poolside a model that developers can inspect, download and operate without committing to Poolside's hosted service. Its commercial test comes next: whether the model is capable enough for teams to move high-volume coding work away from larger proprietary systems, and whether those teams will pay Poolside to manage that infrastructure inside their own security boundaries.