🦙 Llama 3 Is Out. It's The Largest Open-Source LLM

Meta has released two versions of the Llama 3 model: 8B and 70B — with 8 billion and 70 billion parameters, respectively. The models are already integrated into Meta AI's virtual assistant in the US and several English-speaking countries.

💪 The Biggest LLM

Meta's AI Chief Scientist, Yann LeCun , writes that Llama 3's context length is 8,000 tokens (approximately 5,000 words in English) — how much the model can remember in a single conversation with a user. The models are trained on 15 trillion tokens on a specially created cluster of 24,000 GPU, making Llama 3 the world's largest LLM model with free access and open-source code.

🔋 Enhanced Performance

Meta notes that Llama 3 can easily handle multi-step tasks thanks to improved scalability and performance, and post-training refinements significantly improve answer consistency. According to the tests published on the Llama 3 page, both the 8B and 70B models outperform models in their class: Mistral 7B and Claude 3 Sonnet, respectively

Look at the chart above. For example, MMLU stands for language understanding, GSM-8K for mathematics, and HumanEval for programming.

👍 Special Scenarios

Llama 3 supports 12 key scenarios: advice seeking, brainstorming, classification, closed-question answering, coding, creative writing, summarization, empathy, open-question answering, reasoning, paraphrasing, and summarizing.

In the coming months, Meta plans to release new versions of the model, including 400B, which, according to the creators of Llama 3, will mark a new milestone in developing open-source LLM.

✅ How to Use?

You can test the new models using Perplexity and select Llama 3 from the dropdown list.

You can read more on our Telegram channel here: https://t.me/hiaimediaen