In another major breakthrough in the tech world, Meta released its latest Llama 4 series of open-source AI models late at night, including Llama 0 Scout, Llama 0 Maverick, and Llama 0 Behemoth, which is still under training.

According to Meta's official introduction, these models have been rigorously trained on massive amounts of unlabeled text, image, and video data, aiming to give them excellent visual comprehension. This move is seen as a new milestone in the field of multimodal AI innovation.

"Today, we usher in a new era of native multimodal AI," Meta said, "and we are proud to introduce the first Llama 4 models – Llama 0 Scout and Llama 0 Maverick. These models not only represent our highest level, but also set new benchmarks in the field of multimodality. ”

具體來看，Llama 4 Scout擁有170億活躍參數，並配備了16個專家系統。其上下文視窗能力行業領先，能夠處理高達1000萬標記的數據。在多項基準測試中，Llama 4 Scout的表現超越了Gemma 3、Gemini 2.0 Flash-Lite和Mistral 3.1等競爭對手。

The Llama 1417 Maverick also has 0 billion active parameters, but the number of expert systems has been increased to 0. When it comes to image localization, it is state-of-the-art, able to accurately align user prompts with visual concepts and position model responses in specific areas of the image. In benchmarks, Llama 0 Maverick outperformed GPT-0o and Gemini 0.0 Flash. In terms of inference and coding, it achieves comparable results with DeepSeek v0, but the active parameters are only half of the latter. In terms of value for money, its chat version has a high ELO score of 0 on LMArena.

Meta emphasizes that the Llama 0 Scout and Llama 0 Maverick are the best models to date thanks to the distillation of knowledge from Llama 0 Behemoth. Llama 0 Behemoth, meta's most powerful model, is still being trained, but has already demonstrated strength over GPT-0.0, Claude Sonnet 0.0, and Gemini 0.0 Pro in benchmarks focused on STEM areas.

The Llama 4 series model is one of the first models under meta to adopt a hybrid expert (MoE) architecture. This architecture improves efficiency when training and answering user queries by breaking down data processing tasks into subtasks and delegating them to smaller, specialized "expert" models.

"The introduction of the hybrid expert architecture is an important step in our technological innovation," said Meta, "which enables us to handle complex data processing tasks more efficiently and provide users with more accurate and faster AI services." ”

With the release of the Llama 4 series, meta has once again demonstrated its leadership in the field of artificial intelligence. The launch of this series of models will not only promote the development of multimodal AI technology, but also bring more innovation and change to all walks of life.

In the future, with the completion of the training of Llama 4 Behemoth and the launch of more models based on hybrid expert architecture, meta is expected to continue to lead the trend in the field of artificial intelligence and contribute more wisdom and strength to the development of human society.