Poolside launches open Laguna model for offline agentic coding
Poolside, a San Francisco-based AI startup founded in 2023, released two new Laguna language models optimized for agentic coding workflows. Laguna XS.2, a 33-billion parameter open-source model under Apache 2.0 license, can run locally on a single GPU without internet connectivity. The larger Laguna M.1, a 225-billion parameter proprietary model, is available free via API and third-party platforms. Both models were trained from scratch rather than fine-tuned from existing base models, and Poolside emphasizes deployment advantages for government and enterprise customers requiring offline, isolated environments.
TL;DR
- →Poolside launched Laguna XS.2, an open-source 33B parameter MoE model under Apache 2.0 license, designed to run locally on single GPUs for private, offline agentic coding
- →The company also released Laguna M.1, a larger 225B parameter proprietary model, available free temporarily via API and partners like OpenRouter, Ollama, and Baseten
- →Both models were trained from scratch using Poolside's internal 'Model Factory' infrastructure and Muon optimizer, which reportedly accelerates training approximately 15 percent faster than industry standard methods
- →Poolside positions itself as faster to deploy for enterprise and government customers needing on-premises, fully isolated environments that work offline, differentiating from OpenAI, Anthropic, and Google
Why it matters
The release represents a rare U.S.-based open-source alternative to the proprietary model arms race between Anthropic and OpenAI, while competing with lower-cost Chinese models like DeepSeek. Poolside's emphasis on offline deployment and government-grade security addresses a specific market gap where data isolation and on-premises execution are non-negotiable requirements. This signals growing demand for capable open models that don't require cloud connectivity or third-party API dependencies.
Business relevance
For developers and operators, Laguna XS.2 enables local agentic coding without cloud costs or latency, reducing operational friction for teams building autonomous agents. For enterprises and government agencies, Poolside's on-premises deployment model and offline capability eliminate data residency concerns and vendor lock-in, making it a practical alternative to proprietary solutions. The free temporary access to Laguna M.1 via multiple distribution channels lowers barriers to testing and adoption.
Key implications
- →Open-source agentic models are becoming competitive on performance and efficiency, potentially fragmenting the market for proprietary coding assistants and agent frameworks
- →Offline-capable models reduce dependency on cloud infrastructure and API providers, enabling organizations with strict data governance or air-gapped environments to deploy AI agents locally
- →Poolside's government and public sector focus suggests a viable business model for AI startups targeting regulated industries where proprietary U.S. labs face deployment friction
What to watch
Monitor adoption rates of Laguna XS.2 among developers and enterprises, particularly in government and defense sectors where offline deployment is critical. Watch whether Poolside's claimed 15 percent training efficiency gain with the Muon optimizer becomes a replicable competitive advantage or is matched by competitors. Track whether the temporary free access to Laguna M.1 converts to sustained usage and whether Poolside eventually monetizes the open-source model or keeps it free.
vff Briefing
Weekly signal. No noise. Built for founders, operators, and AI-curious professionals.
No spam. Unsubscribe any time.
