Poolside

Poolside (also referred to as Poolside AI) is an artificial intelligence company that trains foundation models from scratch specifically for software engineering, rather than for general-purpose tasks. Its models are exposed through agentic coding products, self-managed inference, an OpenAI-compatible API, and agent workflows aimed at organizations that need to run AI inside infrastructure they control, including virtual private clouds, on-premises data centers, and air-gapped environments. The stated purpose of this deployment model is to keep customer code, prompts and data within the customer's own security boundary while providing predictable inference costs and auditability.

The company's current model line is the Laguna family, released with public weights. As described on the company website, Laguna S 2.1 is a 118B-parameter model with 8B active parameters and a 1M-token context window positioned for reasoning at mid-size cost, and Laguna XS 2.1 is a 33B-parameter model with 3B active parameters and a 256K context window small enough to run on-device. Earlier Laguna XS.2 and M.1 models were released in April 2026 together with the runtime used to train and operate agents. Product surfaces include the Poolside Agent CLI (pool and pool exec), tool integrations with ACP-compatible editors, a macOS Desktop Assistant, and extensions for VS Code and Visual Studio. Capabilities described in Poolside's documentation include code generation and completion with local context awareness, explaining and refactoring legacy code, test generation, documentation, incident investigation, security reviews, and multi-step task automation, with approval prompts, tool permissions and sandboxing to constrain agent actions.

Poolside also publishes research and engineering write-ups, including a six-part "Model Factory" series covering its training and evaluation infrastructure (data pipelines, the Titan distributed training codebase, code execution environments, inference and evaluation, and post-training at scale), plus work on reinforcement learning from code execution feedback, benchmark and reward hacking, and NVLink C2C activation offloading on NVIDIA Grace Blackwell superchips. Beyond software, the company has pursued vertical integration into compute: Project Horizon, announced in October 2025 with CoreWeave, is a 2-gigawatt AI campus in Pecos County, West Texas.