Poolside

From an enterprise Code LLM pitch to an open-weight model lineup

Poolside began as a company built around a distinctly enterprise pitch: instead of asking large software organizations to send their proprietary source code to a third-party cloud-based coding assistant, it offered a foundational programming language model — originally referred to as the Malibu family — that could be trained or fine-tuned on a company’s own internal codebase and deployed inside that company’s private cloud or on-premise environment. The companion product, called Assistant, plugged into everyday IDEs like VS Code and JetBrains so developers could keep working in familiar tools while the underlying model stayed private.

The current version of Poolside’s own site shows the company has since pushed further into open-weight territory. It now promotes a newer model family called Laguna, offered in at least two sizes: Laguna xs 2.1 (33B parameters, small enough to run on-device, 256K context) and Laguna s 2.1 (118B parameters with 8B active, 1M context), described as “frontier-class reasoning at mid-size cost.” These are positioned as open-weight, agentic coding models, and Poolside explicitly points developers to run them through a desktop app, a CLI, OpenRouter, or the Vercel AI Gateway — a much more accessible distribution model than the earlier private-deployment-only narrative.

pool: the command-line agent at the center of the workflow

The centerpiece of the current product experience is a CLI/agent tool named pool (version 1.0.6 at the time this was reviewed). It behaves like an agentic terminal assistant: it connects to an agent server, lets you switch between models (Laguna s 2.1 and others) mid-session, group tool calls, open prompts in an external editor, and run slash commands. It supports project-aware conversations — for example scoping a chat to a specific repo and branch (the interface shows examples like a “feature/api-retries” branch tied to a pull request) — and keeps separate chat histories per project, such as a website repo, a model-runtime repo, or documentation. Typical tasks shown in the interface include planning a safe API migration, summarizing benchmark results, implementing a retry mechanism for failed API calls, profiling model latency, and refreshing CLI documentation — in other words, real engineering chores rather than toy autocomplete demos.

Open research posture

Poolside publishes a recurring “Model Factory” blog series describing how it trains, evaluates, and iterates on its foundation models — covering topics like GPU-to-GPU weight transfers, automated architecture ablations, reinforcement learning from code execution at scale, reward hacking in benchmarks, and low-level infrastructure work such as using NVIDIA’s NVLink C2C on Grace-based superchips to offload MLP activations during training for measurable throughput gains. This is unusually technical, systems-level content for a product website, and it signals that Poolside is trying to build credibility with ML infrastructure engineers and researchers, not just enterprise buyers.

Government and mission-driven framing

Alongside the developer-facing tooling, Poolside maintains a dedicated Government offering and talks explicitly about bringing “frontier AI to your mission,” plus a broader stated vision that “the fastest path to AGI runs through software.” This positions the company as aiming well beyond a simple coding-assistant plugin, toward being a foundation-model vendor for both software engineering and higher-stakes public-sector use cases.

Who this actually fits

• Enterprises with strict source-code confidentiality requirements that previously ruled out public coding assistants may still find the original private-deployment angle relevant, though the site’s current emphasis is on open-weight models distributed via CLI, desktop app, and third-party gateways rather than a packaged on-prem product.
• ML infrastructure teams and researchers who want transparency into training methodology will get more value from the Model Factory blog than a typical end-user developer will.
• Individual hobbyist developers looking for a lightweight VS Code autocomplete plugin are not really the target audience here — the tooling leans toward agentic, project-scoped, terminal-driven workflows and government/enterprise deployments.
• Teams needing clear, published self-serve pricing should be cautious: the site text reviewed does not display pricing tables, so anyone evaluating Poolside for budget purposes will need to contact the company directly.

Overall, Poolside reads less like a finished SaaS product and more like a fast-moving foundation-model lab (with a research-preview flavor around its newest Laguna models) that also ships a real, usable CLI agent. Prospective users should check the official site directly for the latest model versions and access options, since this space is evolving quickly.