On April 5, Meta Platforms announced the launch of a new model of its artificial intelligence, Llama 4. The new model is already available to users in two versions: Llama 4 Scout and Llama 4 Maverick. The company announced the launch in its blog.
What is new in Llama 4?
According to Meta, Llama is a multimodal system that can process, integrate and convert between different types of data, such as text, video, images and audio.
The new model will be available in three versions:
- Lama 4 Scout is a multimodal model with 17 billion active parameters and 16 experts. According to the company, it is the best in its class, more powerful than previous Llama models, but fits on a single H100 GPU. Scout has a contextual window of 10 million tokens and is said to outperform Gemma 3, Gemini 2.0 Flash-Lite and Mistral 3.1 in many tests.
- Llama 4 Maverick - also has 17 billion active parameters, but uses 128 experts . Meta says the model outperforms GPT-4o and Gemini 2.0 Flash in many tests, and gives similar results to DeepSeek v3 in inference and coding with half the parameters.
- Both models are based on Llama 4 Behemoth, a large model with 288 billion parameters and 16 experts. This is the most powerful meta model to date, outperforming GPT-4.5, Claude Sonnet 3.7, and Gemini 2.0 Pro in tests. Behemoth is still in the training phase, but the company is ready to share the results in previews.
AIN reminds that Meta AI recently launched its AI chatbot in the EU and several other countries, but with some restrictions. This was the largest geographic global expansion of its AI for the company.
The new launch comes amid a stock market crash for tech companies in the US, including Meta. The day before the announcement, it was revealed that Zuckerberg and other billionaires lost nearly $208 billion in one day due to Donald Trump's new tariffs.
