San Francisco-based AI startup Poolside announced the release of two new artificial intelligence models designed for software development tasks, including one available under an open-source license. The company, founded in 2023, launched Laguna M.1, a proprietary 225-billion parameter model, and Laguna XS.2, a 33-billion parameter model released under the Apache 2.0 open-source license.

Poolside's models are optimized for what the company calls "agentic workflows" - AI systems that can write code, use third-party tools, and take autonomous actions beyond simple text generation. The larger M.1 model targets enterprise and government environments for complex software engineering problems, while the smaller XS.2 is designed to run locally on consumer hardware including laptops and desktop computers.

Benchmark testing showed Laguna M.1 achieved a 46.9% score on SWE-bench Pro, a standard for measuring AI performance on real-world software issues, while the smaller XS.2 scored 44.5% despite having significantly fewer parameters. Both models can run without internet connectivity, addressing security concerns for government and enterprise users who require offline capabilities.

Separately, researchers from JD.com and academic institutions introduced a new training technique called Reinforcement Learning with Verifiable Rewards with Self-Distillation (RLSD). This method aims to reduce the computational costs of training AI reasoning models by combining elements of reinforcement learning with self-distillation techniques.

The RLSD approach addresses limitations in current training methods, which either provide sparse feedback through trial-and-error learning or require expensive teacher models for guidance. In testing on the Qwen3-VL-8B vision-language model, RLSD achieved 56.18% average accuracy across visual reasoning benchmarks, outperforming standard training methods while requiring roughly half the training time.

Poolside has also released supporting tools including "pool," a terminal-based coding agent, and "shimmer," a web-based development environment that can run on mobile devices. The company stated its goal of supporting the broader AI research community while maintaining its focus on government and enterprise clients who require secure, deployable AI systems.