News

Enterprise

Artificial Intelligence

Americas

Meta Unleashes Llama 4: A New Era of Open Multimodal AI Innovation

Meta has launched the Llama 4 herd, introducing natively multimodal open-weight models that challenge proprietary industry leaders.

Meta has launched the Llama 4 herd, introducing natively multimodal open-weight models that challenge proprietary industry leaders.

NewDecoded

Published Apr 7, 2026

Apr 7, 2026

3 min read

Image by Meta

Meta officially released its Llama 4 model family today, marking a significant shift toward open source multimodal intelligence. The collection, known as the herd, includes the efficient Llama 4 Scout and the powerhouse Llama 4 Maverick. These models are the first from Meta to utilize a Mixture of Experts architecture and native multimodality for processing text, images, and video simultaneously.

Llama 4 Scout introduces an industry leading 10 million token context window, allowing it to reason over massive codebases or document sets. Despite its performance, it fits on a single NVIDIA H100 GPU using standard quantization. This efficiency positions it as a superior alternative to compact models like Google’s Gemini 2.0 Flash Lite.

The larger Llama 4 Maverick model outperforms GPT-4o and Gemini 2.0 Flash in core benchmarks. It achieves these results with fewer active parameters than its competitors, offering an attractive performance to cost ratio. Testing shows that Maverick matches the reasoning skills of DeepSeek v3 while remaining more efficient to deploy.

Behind these releases stands Llama 4 Behemoth, a massive 2 trillion parameter internal teacher model. Behemoth surpasses frontier systems such as GPT-4.5 and Claude 3.7 Sonnet on STEM and math evaluations. Meta used co distillation to transfer this advanced intelligence into the smaller, publicly available Scout and Maverick models.

Unlike previous generations, Llama 4 uses early fusion to integrate text, images, and video natively. This breakthrough allows the models to understand spatial grounding in images within the same transformer backbone. Such integration ensures the AI interacts naturally with visual inputs without losing linguistic nuance.

The release addresses industry concerns regarding bias and refusal rates. Meta reduced the refusal rate on debated topics to below 2 percent, down from 7 percent in Llama 3.3. New safety tools like Llama Guard and the GOAT red teaming system were also introduced to support secure development.

Llama 4 Scout and Maverick are available for download on Hugging Face and llama.com. Consumers can also access the new intelligence through Meta AI on WhatsApp and Instagram platforms today.


Decoded Take

Decoded Take

Decoded Take

The release of Llama 4 signals a definitive end to the era where proprietary models held a monopoly on frontier multimodal capabilities. By combining a 10 million token context window with native video processing, Meta is forcing competitors to rethink their restrictive access models and high pricing. This release effectively democratizes high end AI development, moving the industry toward a future where sophisticated reasoning and massive data ingestion are no longer restricted to closed systems.

Share this article

Related Articles