Poolside

Free

Poolside builds open-weight AI models trained specifically for software development. You run them in your own environment, not someone else's cloud.

The company offers two core models: Laguna S 2.1, a 118 billion parameter model with one million token context, and Laguna XS 2.1, a smaller 33 billion parameter version that runs on-device. Both are built for agentic coding tasks.

You access them through a desktop app, command line, or integrations like OpenRouter and AWS Bedrock. This is coding assistants work, with a deliberate focus on teams who cannot ship code to third-party APIs.

Open Poolside if you control your own infrastructure and need code to stay inside it. Skip it if you want something that works out of the box without setup.

The fit is narrow but deep for these situations.

• Enterprise engineering teams with strict data residency requirements: Models run inside your VPC. Your code never leaves your infrastructure.
• Government agencies and defense contractors: Air-gapped deployment is supported. This is built for classified or regulated environments.
• Organizations with large, complex codebases where context windows matter: The one million token context fits entire repositories for refactoring tasks.
• Teams that need fine-tuned models on their specific architectures: You get the full weights to adapt to your code patterns.
• Users who prioritize backing technically strong founding teams: The founders shipped Copilot at GitHub and built developer platforms before.
• Large enterprises with existing accelerator infrastructure: You already have GPUs warm. Poolside amortizes across hardware you own.
• Buyers who prioritize model-and-delivery over application-layer portability: You want the engine, not a pre-built app.
• Developers wanting terminal-first agent interaction: The pool CLI is designed for keyboard-driven workflows.
• Security-conscious users wanting granular permission control: Approval modes range from fully manual to risk-based auto, not just binary allow/block.

These are what you notice after running it yourself.

• RLCEF training on actual execution: Models learn from whether code compiles and tests pass, not just pattern matching. You get fewer broken suggestions.
• Complete benchmark transparency: Poolside publishes unedited trajectories including failures and reward hacking attempts. You can see exactly how the model behaves.
• Thoughtful permission system: Risk-based auto mode distinguishes low and high stakes operations. You are not asked to approve every keystroke.
• Build and Plan agent modes: Plan mode lets you inspect before the agent acts. You switch to Build when you are ready to let it run.
• No per-token cloud fees on owned hardware: Once you have the weights, inference cost is electricity and depreciation. At scale, this changes the math.

You can run it a few ways depending on your stack.

1. Install the pool CLI or desktop app for terminal-first interaction.
2. Connect via AWS Bedrock if you need first-party procurement and unified billing.
3. Route through OpenRouter or Vercel AI Gateway for flexible access.
4. Configure approval modes with Shift+Tab to set your risk tolerance per session.
5. Run pool exec for non-interactive CI or automation tasks.

These limits trip up most shoppers.

• Heavy setup and procurement: You sign custom contracts, review legal terms, and install on your own hardware. This is not a signup-and-go tool.
• Laguna S fails at structured JSON: The larger model cannot reliably return valid JSON for agent tasks, despite the context window. You may end up using the smaller XS model instead.
• Slower than cloud alternatives: Laguna XS runs reliably but needs 3x more calls and 3.8x more tokens than Groq for the same work. You trade speed for cost and control.
• No image input: Neither model accepts screenshots. You need a second provider for visual validation tasks.
• Leadership execution concerns: The engineering team is strong but management lacks experience, and the company is spread across too many initiatives for a young foundation model startup.