Meta Releases Llama 4 Openly — and It Outpaces GPT-4o on Code

Meta released Llama 4 this week under a permissive open weights licence, and the timing feels deliberate. The Llama 4 family ships in three sizes: Scout (17B parameters, MoE architecture), Maverick (400B MoE), and a not-yet-released Behemoth.

The MoE architecture is the key engineering story: by activating only a fraction of parameters per forward pass, Meta achieves GPT-4-class performance at a fraction of the inference cost.

The competitive implications for enterprise AI are significant. Any organisation currently evaluating proprietary models for internal deployment now has a credible open alternative that can be run on-premises, fine-tuned on proprietary data, and served without per-token API costs. That changes the build-versus-buy calculus considerably.