Recently, Meta announced the release of Llama 4, its next-generation suite of open-weight, multimodal AI models designed to power more personalized and intelligent experiences. The first two models in the series which include Llama 4 Scout and Llama 4 Maverick are now available for download, while a more advanced model, Llama 4 Behemoth and Llama 4 Reasoning , remains in training.

Introducing Llama 4 Scout and Llama 4 Maverick

Built on a mixture-of-experts (MoE) architecture, these models balance efficiency and performance:

1. Llama 4 Scout (17B active parameters, 16 experts) is optimized for efficiency, extremely fast, fitting on a single NVIDIA H100 GPU while supporting an industry-leading 10 million token context window. It outperforms comparable models like Gemma 3 and Mistral 3.1 in reasoning, coding, and multimodal tasks.
2. Llama 4 Maverick (17B active parameters, 128 experts) rivals GPT-4o and Gemini 2.0 Flash in benchmarks, it's much smaller and efficiency than DeepSeek 3, delivering strong performance in reasoning, coding, and image understanding at a fraction of the computational cost.

Both models are natively multimodal, integrating text and vision seamlessly through early fusion, enabling richer interactions with images and videos.

The Power Behind Llama 4: Llama 4 Behemoth

The smaller models were distilled from Llama 4 Behemoth, a 288B-parameter MoE model still in training. Early benchmarks show it surpassing GPT-4.5, Claude Sonnet 3.7, and Gemini 2.0 Pro in STEM-focused tasks. Meta plans to share more details as training progresses.

Key Innovations

1. Efficiency & Scalability: MoE architecture ensures only a subset of parameters are activated per inference, reducing costs.
2. Extended Context: Llama 4 Scout’s 10M-token support enables deep analysis of long documents, codebases, and user activity logs.
3. Improved Training Techniques: Meta introduced MetaP, a method for optimizing hyperparameters, and refined reinforcement learning to enhance reasoning and coding abilities.
4. Bias Mitigation: Llama 4 shows reduced political bias compared to Llama 3, with refusal rates on debated topics dropping from 7% to under 2%.

Safety & Accessibility

Meta continues its commitment to open AI development, releasing safeguards like:

1. Llama Guard (input/output safety filtering)
2. Prompt Guard (jailbreak detection)
3. CyberSecEval (cybersecurity risk assessment)

Developers can download Llama 4 Scout and Maverick today on llama.com and Hugging Face, with integrations rolling out across WhatsApp, Messenger, Instagram, and Meta.AI.

Download the models and explore the future of AI.