Meta Releases Llama 4 Open-Source — Rivals Closed Models on Coding Tasks

Meta released the Llama 4 model family under a permissive open licence, making available four model sizes including a 400 billion parameter Mixture-of-Experts (MoE) variant that rivals leading closed-source models on coding and reasoning benchmarks. The models are available on Hugging Face and via Meta's AI Studio.

Llama 4 Scout (17B active parameters from a 109B MoE) and Llama 4 Maverick (17B active from 400B MoE) represent particularly interesting value propositions — near frontier quality at significantly lower inference cost due to the sparse activation of MoE architectures.

Model family breakdown

• —Llama 4 Scout: 17B active / 109B total — lightweight, fast, for edge and mobile
• —Llama 4 Maverick: 17B active / 400B total — frontier quality at MoE efficiency
• —Llama 4 Behemoth: 288B active / 2T total — research preview, not yet released
• —Context window: 10M tokens on Scout and Maverick
• —Licence: permissive commercial use for most use cases

Open source is winning the intelligence race. Llama 4 Maverick outperforms GPT-4o and Gemini 2.0 Flash on a majority of benchmarks at a fraction of the cost. — Mark Zuckerberg