Meta plans to release Llama 3 in July, with capabilities close to GPT 4 and up to 140 billion parame

Mondo Digital Updated on 2024-02-29

On Wednesday, local time, the information of science and technology **the information quoted people familiar with the matter reported,Meta plans to release the LLAMA 3 model in July this year.

The mission of Llama 3 is enormous.

According to the reportMeta hopes that LLAMA 3 can compete with OpenAI's GPT-4, which has become a powerful multimodal model capable of not only processing longer texts, but also supporting image input.

But a Meta employee revealed that because researchers have not yet begun fine-tuning the model,Therefore, the company has not yet decided whether LLAMA 3 is multimodal or not. Fine-tuning is the process by which developers provide additional data to an existing model so that a large model can learn new information or work on a task.

Meta employees also saidLlama 3 is likely to have a maximum of more than 140 billion parameters, compared to the maximum of 70 billion parameters for Llama 2, which was released last July.

As previously reportedThe GPT-4 model parameter scale is about 18 trillion, LLAMA 3 is less than a tenth of that number.

In addition, before the launch of LLAMA 3, Meta was still trying to overcome a problem found in LLAMA 2 - not being able to deal with any controversial issues.

Due to the fact that the developers added safety guardrails to Llama 2, it refuses to answer a series of questions that are considered controversial.

According to Meta employees, the guardrails make Llama 2 appear "too safe" in the eyes of the company's top leaders and model researchersThe researchers plan to ease the restrictions of LLAMA 3 in this regardso that it engages more with the user, providing context rather than just refusing to answer.

Expectations for LLAMA 3 are growing, but Meta still faces a protracted competition for talent.

Louis Martin, a researcher responsible for the security of Llama 2 and Llama 3, left the company this month, according to two people familiar with the matter. Kevin Stone, the head of reinforcement learning, also left this month, according to one of the people.

Wall Street news, welcome **app to see more.

Related Pages