Three open-weight models shipped between April 2025 and February 2026—Meta’s Llama 4 Maverick, Mistral AI’s Large 3, and MiniMax’s M2.5—offer downloadable weights for self-hosting on GPUs, a key consideration for Irish and EU teams facing data residency rules, the EU AI Act’s risk-based obligations, and a push for vendor independence. Meta released Llama 4 Maverick on April 5, 2025; Mistral AI, the Paris-based lab, followed on December 2, 2025 with Mistral Large 3; and MiniMax, a Chinese AI company, shipped M2.5 on February 12, 2026. The comparison covers specs, pricing, published benchmarks, and licensing terms, then walks through real deployment scenarios, a migration guide, and a verdict built on the numbers rather than launch-day marketing.