Meta launches Llama 4: A new era of Open Multimodal AI

Meta has announced Llama 4, its most powerful AI model to date. Llama 4 is a new generation of natively multimodal, open-weight, and high-context large language models. The launch includes three major models: Llama 4 Scout, Llama 4 Maverick, and the yet-to-be-released Llama 4 Behemoth, which push the frontier of general-purpose AI across reasoning, image, understanding, and code generation.

A Leap in MultiModal Intelligence

One of the standout features of the Llama 4 family is their native multimodal capability. Unlike previous generations of Llama that bolted on image understanding post-training or through adapters, Llama 4 models were jointly trained on text, image, and video data using a deeply integrated architecture.

In terms of architectural brakthroughs, Meta has also introduced Mixture-of-Experts (MoE) models for the first time in the Llama series. Instead of activating the entire model for every input, these models selectively activate a small subset of specialized experts per token, improving performance without the usual compute costs.

The context window size has also been dramatically extended. Llama 4 Scout supports up to 10 million tokens, which is the longest context window of any openly available model to date. This opens up new possibilities for analyzing entire books, codebases, research corpora, or even multimedia archives in a single pass. The architecture supporting this, known as iRoPE (infinite RoPE), allows Scout to scale context dynamically without retraining.

These context advances also benefit multimodal tasks. For instance, Scout has demonstrated top-tier performance in summarizing long image-rich documents, maintaining coherence and factuality over thousands of interleaved text-image inputs.

Performance and Efficiency Across the Board

Each model serves a distinct use case. Scout, with 17 billion active parameters, is designed for efficiency and portability, small enough to run on a single H100 GPU but powerful enough to outperform many dense models on multi-document reasoning. Maverick, also with 17B active parameters but backed by a 400B MoE architecture, offers the best performance-to-cost ratio among open models and excels at assistant-style tasks. According to Meta, Maverick scores a 1417 ELO on LMArena, making it competitive with models like GPT-4o and Claude 3 Sonnet.

At the top of the stack, Llama 4 Behemoth is still in the oven but already making waves as Meta’s new frontier model. It boasts 288 billion active parameters and a staggering 2 trillion total, including a deeper, asynchronous RL pipeline and next-gen MoE layers. Behemoth already outperforms GPT-4.5 on STEM benchmarks and serves as the training teacher for Scout and Maverick.

Safety, Guardrails, and Open Access

As with previous Llama models, Meta is committed to safe and responsible AI development. New tools and evaluations for safety accompany the release of Llama 4. Llama Guard provides robust input and output moderation, Prompt Guard detects and mitigates prompt injections and jailbreaks, and CyberSecEval benchmarks model resilience against security exploits. An innovative system called GOAT (Generative Offensive Agent Testing) uses automated adversarial probing to simulate real-world misuse scenarios.

Despite their advanced capabilities, Scout and Maverick are freely available under Meta’s open license. Researchers, developers, and enterprises can download the models from Llama official page or Hugging Face, reaffirming Meta’s position as a champion of open AI development.

Looking Ahead

With Llama 4, Meta isn’t just catching up with proprietary models, it’s reshaping the open-source AI ecosystem. These models are fast, capable, and now offer native multimodal reasoning at scale, setting a new bar for research, application development, and even academic benchmarking.

Read More:

The Speed of Money Demands Smarter Machines

Linedata Acquires AI Specialist NROAD

India to unveil Indigenous Wireless EV Charger for EVs, Railways, and Green Buildings

Intel Core Ultra speeds routine tasks by 50% for Deloitte’s AI