Meta launches Llama 4 models; super voracious, candid as Grok

Meta on Saturday unveiled the Llama 4 series — its most advanced collection of open-weight large language models (LLMs) to date. 

The release includes Llama 4 Scout, Llama 4 Maverick, and a preview of the much larger Llama 4 Behemoth. 

According to Meta, the new Llama 4 models outperform previous generations and rival even the most powerful proprietary systems in existence. 

Unlike closed models such as Grok or OpenAI's GPT-4, Llama 4 models are being made available for download and deployment via Meta's platforms and through Hugging Face, enabling researchers and developers to build custom applications.

The Instagram-owner framed the release as a milestone in its vision for open AI development. "We continue to believe that openness drives innovation and is good for developers, good for Meta, and good for the world." 

According to a release, Llama 4 Scout and Maverick are both 17-billion-parameter models, but differ in design. Scout leverages 16 experts and is optimised for low-latency, cost-effective deployment—even fitting on a single NVIDIA H100 GPU with quantisation. 

Meanwhile, Maverick features a Mixture-of-Experts (MoE) architecture with 128 experts, enabling it to tap into 400 billion parameters selectively, boosting inference efficiency and reducing costs.

Unprecedented Context Lengths and Multimodal Capabilities

A standout feature of Llama 4 Scout is its 10 million-token context window—an industry-leading leap that allows multi-document summarisation and reasoning across vast codebases. 

In comparison, even the most advanced models such as GPT-4 and Grok support far smaller contexts. 

Meta says both Scout and Maverick are natively multimodal, supporting seamless interaction with text, images, and video. 

According to the company, this capability makes Maverick especially well-suited for image understanding and creative writing—critical for next-generation digital assistants and AI-powered content generation.

While not yet publicly available, Meta previewed Llama 4 Behemoth, a teacher model used to guide the training of Scout and Maverick. 

With 288 billion active parameters, 16 experts, and nearly 2 trillion total parameters, Behemoth reportedly outperforms top-tier models such as GPT-4.5, Claude Sonnet 3.7, and Gemini 2.0 Pro on demanding benchmarks like MATH-500 and GPQA Diamond.

Meta said that Behemoth was instrumental in codistillation, a training process that allowed the more compact Scout and Maverick models to inherit advanced capabilities while keeping computational requirements manageable.

Competing with Grok on Political Neutrality and Candour

Beyond technical breakthroughs, Meta is positioning Llama 4 as a more balanced and candid model in handling politically and socially contentious topics — a space where many leading LLMs struggle. 

Compared to Llama 3, the latest generation shows a dramatic reduction in refusal rates when responding to debated topics, falling from 7 per cent to under 2 per cent, and showing improved balance in which questions it declines to answer.

Meta claims that Llama 4 now performs at parity with Grok, Elon Musk's xAI model, in terms of political lean and refusal patterns. Notably, the rate at which Llama 4 shows strong political bias is half that of Llama 3.3, suggesting meaningful progress in the effort to eliminate model bias.

"Llama 4 is dramatically more balanced," the company stated, emphasising its ability to present multiple sides of contentious issues without judgment or preference.

Meta is set to unveil further details about the Llama 4 ecosystem at LlamaCon on April 29, where it will discuss its research roadmap and future model iterations. 

Image Source: Unsplash