The artificial intelligence landscape has witnessed an intriguing development as American startup Poolside releases two groundbreaking large language models designed specifically for agentic coding workflows. Whilst tech giants Anthropic and OpenAI continue their expensive model rivalry, and Chinese companies like DeepSeek pursue affordable alternatives, this San Francisco-based company founded in 2023 has emerged with a compelling proposition: high-performance AI models that are both accessible and, in one case, completely open-source.
Poolside's release comprises two distinct models. The flagship Laguna M.1 is a proprietary 225-billion parameter Mixture of Experts model with 23 billion active parameters, optimised for complex enterprise and government software engineering challenges. However, the more disruptive offering is Laguna XS.2, a 33-billion parameter model with just 3 billion active parameters, released under the permissive Apache 2.0 open-source licence. This smaller model can run entirely offline on consumer hardware—requiring as little as 36GB of unified memory on Apple Silicon or 24-32GB VRAM on PCs with quantisation—delivering complete privacy and security for developers working on sensitive projects.
What makes these models remarkable is their performance relative to size. On the SWE-bench Pro benchmark, which tests real-world software problem-solving abilities, Laguna M.1 achieved 46.9% whilst the considerably smaller XS.2 scored an impressive 44.5%, nearly matching its larger sibling and outperforming competitors like Claude Haiku 4.5 and Gemma 4. These models weren't simply fine-tuned from existing frameworks but trained from scratch using Poolside's proprietary "Model Factory" infrastructure, incorporating 30 trillion tokens of carefully curated data—13% of which is synthetic, custom-made practice material designed to teach specific skills.
Poolside's training methodology employs sophisticated systems including the Muon optimiser, which accelerates learning approximately 15% faster than standard methods, and AutoMixer, which scientifically determines optimal combinations of code, mathematics, and general knowledge. Following initial training, models undergo reinforcement learning in virtual environments where they practice solving real software engineering problems through trial and error, receiving rewards for successful bug fixes and working code—transforming them from text generators into capable autonomous agents.
Beyond the models themselves, Poolside has introduced two complementary tools. "Pool" is a terminal-based coding agent that functions as an Agent Client Protocol server, bringing researchers' internal training tools to the public. "Shimmer" is a cloud-native development environment offering instant-on virtual machine sandboxes with integrated AI agents, remarkably designed to function even on smartphones—suggesting a future where complex engineering work isn't confined to desktop computers. The company's decision to release XS.2 under Apache 2.0 licensing, allowing unrestricted commercial use without royalties, reflects their conviction that "the West needs strong open-weight models" and positions Poolside as a cornerstone of the open-AI ecosystem, directly challenging the predominantly closed approaches of major competitors.
Fuente Original: https://venturebeat.com/technology/american-ai-startup-poolside-launches-free-high-performing-open-model-laguna-xs-2-for-local-agentic-coding
Artículos relacionados de LaRebelión:
- Xiaomis Open-Source AI Models Dominate Agentic Tasks
- Colorado Protects Open Source in New Age Verification Law
- DeepSeek V4 Nueva Revolucion en IA Open-Source
- OpenAI Unveils GPT-55 Smarter Coding and Efficiency
- OpenAIs Privacy Filter Open-Source Data Protection
Artículo generado mediante LaRebelionBOT
No hay comentarios:
Publicar un comentario