Meta releases first open-weight Llama 4 family under Apache 2.0, three models for agent execution
Meta released its first Apache 2.0 open-weight models since Llama, with three variants aimed at agent execution across cloud and edge.
On August 18, 2026, Meta released its first three open-weight models under the Apache 2.0 license since the Llama family. The launch placed agent execution at the core. The lineup covers flagship, mid, and lightweight tiers: the flagship targets complex reasoning, multi-step agents, and code generation; the mid-tier targets enterprise applications, RAG, and tool calling; and the lightweight targets edge, on-device, and low-latency scenarios.
Meta strengthened four agent dimensions: tool-calling stability, multi-step task planning, code-OS interaction, and long-context task tracking. Multi-turn function calls, parameter extraction, and result integration error rates declined materially, and the models retained completed steps and current goals across long agent trajectories.
The Apache 2.0 license reduces legal review cost for commercial deployment. Enterprises can run inference, fine-tune, and develop downstream on their own hardware. Downstream teams can prototype with the lightweight tier and scale to the flagship without switching model families.