D. Jassim Haji
Llama4 is a new set of artificial intelligence models that will make “AI” more intelligent on applications such as “WhatsApp”, “Instagram” and “Messenger”. Two models of this series, Llama 4 Scout and Llama4 MAVERICK, can be downloaded from the Llama or via the “Huging Face” platform. The company also reviewed a special model called Llama4 Beheemoth, an advanced model used in training the rest of the Llama4 models.
Simply put, the models available to the public and trained are known as open weights as “artificial intelligence models with open weights.” These models allow users to operate locally without the need to rely on cloud application programming facades. However, there may remain licensing restrictions related to modification or commercial use.
According to reports, Meta designed the Llama4 Scout model, a smaller model, to be in original multimedia. This model depends on the design of a “mix of experts”, which is more like a team of 16 specialized experts working together, with 17 billion active teachers that make it a very smart model. This model can process a huge amount of information in one go of up to 10 million data units, whether it is words or parts of the images. In addition, it has a high efficiency that makes it able to work with only one NVIDIA H100 graphics processing unit, a computer slice specifically designed for heavy tasks in the field of artificial intelligence.
Llama 4 may be the best mathematics teacher for Llama4 Beheemoth the largest and strongest model in the Llama4 family, as it works as a teacher who helps in training smaller models. It is also a multimedia model that can deal with texts, images and others.
This model contains 288 billion active teachers, and the total number of its features reaches a trillion parameter. “Mita” claims that this model has super performance in areas such as mathematics, work in various languages, and image processing, which makes it ideal for directing and training the rest of the Llama4 models.
Moreover, Mita stated that her new artificial intelligence models are able to outperform models such as GEMMA 3 from Google, Gemini2.0 Flash-Lite, Chatgpt-4O of «Openai) and Mistral 3.1 in many tests that measure the performance of artificial intelligence models.
Scout and MAVERICK: Meta presented the Llama4 Scout model, which is a strong and lightweight model at the same time, built using 17 billion active teachers with 16 experts. Scout works efficiently on one advanced graphics card – NVIDIA H100 – which makes it more available to developers and researchers.
In artificial intelligence terms, “parameters” such as nerve cells in the brain work, as the model helps to learn from data, understand the language, and make information based on information. The higher the number of parameters, the higher the degree of intelligence of the model.
SCOUT is also able to process 10 million (TOKEN) icons at once, which makes it very useful in analyzing huge amounts of data and information simultaneously. According to the company in a post on its blog, Scout outperforms previous versions of Llama models and other similar models such as GEMMA3, Gemini2.0 Flash-Lite and Mistral 3.1 in many standard tests.
Compared with China: China is witnessing remarkable progress in its capabilities to develop artificial intelligence, with the launch of the “QWEN” series of “Ali Baba”, “Deepseek’s R1”, “Manusai”, and “Hunyuan Turbo S” from “Tenning”, among other models. The development of the low -cost model in China has been completed in just two months, with an investment less than 6 Millions of dollars, which is a major difference compared to the cost of training the GPT-4 model from «Openai», which amounted to about $ 100 million, according to a report of «ET» in January.
The QWEN model achieves a competitive performance with the leading models and outperforms “Deepseek V3” in the popular programming tests and user information. As for “QWEN 2.5 Max”, it includes 72 billion teachers, and perhaps more in its maximum issuance “Max”, and it was trained in 20 trillion symbol, surpassing “Deepseek”, which trained on 14.8 trillion symbol.
Last month, another Chinese model called Manus AI, which represents a qualitative leap towards independence in artificial intelligence, was presented, where it can carry out a multi -step workflow, and reach reliable data sources through API programming facades. Manus has achieved a global pioneering performance level across all its three levels of difficulty. However, according to a survey conducted by the “ARATIFICIAL Analysis” platform, an independent platform for assessing artificial intelligence performance, the “Llama” model came second as the most common models that are used, and ranked between open source models.
With the expansion of the capabilities of artificial intelligence globally, the industry is increasing towards criteria that are not limited to performance, but also include openness, efficiency, and complementarity between languages and multimedia – areas in which both “dead” and Chinese giants are noticeable.
