LLaMA 3.1 hits the web: ChatGPT has a new competitor with many advantages

A new 405B LLaMa 3.1 has hit the internet. It could quickly become the new king of chatbots, outperforming GPT-4o in most benchmarks.

Tech Jul 27, 2024 0 43 Add to Reading List

LLaMA 3.1 hits the web: ChatGPT has a new competitor with many advantages

The traditional model appeared on the Internet as a torrent file on Reddit before the official release. The 8V and 70V versions were also updated, after which the models improved almost 2 times in some tests.

As benchmarks show, LLaMa 3.1 outperforms GPT-4o on several tests, including GSM8K, Hellaswag, MMLU-humanities, MMLU-other, MMLU-stem, and winograd. However, LLaMa 3.1 lags behind on HumanEval and MMLU-social sciences .

The evaluation was performed by analyzing the base model LLaMa 3.1. To fully realize its potential, the instructions need to be tuned. Experts believe that many results can be significantly improved when the Instruct version of Llama 3.1 models is released. In addition to the quality improvements, this model has increased the context size to 128K from 8K. Time will tell if LLaMa 3.1 can outperform ChatGPT-4o.

Meta* Corporation has previously stated that the next-generation neural network Llama 3 is the most capable of the open LLMs currently available. A few months ago, two versions were released: Llama 3 8B and Llama 3 70B.