Meta says Llama 3 beats most other models, including Gemini

Meta says Llama 3 beats most other models, including Gemini

The next generation of Meta’s large language model Llama, which releases today to cloud providers like AWS and to model libraries like Hugging Face soon, performs better than most current AI models, the company said in a blog post.

Llama 3 currently features two model weights, with 8B and 70B parameters. (The B is for billions and represents how complex a model is and how much of its training it understands.) It only offers text-based responses so far, but Meta says these are “a major leap” over the previous version. Llama 3 showed more diversity in answering prompts, had fewer false refusals where it declined to respond to questions, and could reason better. Meta also says Llama 3 understands more instructions and writes better code than before.

In the post, Meta claims both sizes of Llama 3 beat similarly sized models like Google’s Gemma and Gemini, Mistral 7B, and Anthropic’s Claude 3 in certain benchmarking tests. In the MMLU benchmark, which typically measures general knowledge, Llama 3 8B performed significantly better than both Gemma 7B and Mistral 7B, while Llama 3 70B slightly edged Gemini Pro 1.5.

(It is perhaps notable that Meta’s 2,700-word post does not mention GPT-4, OpenAI’s flagship model.)

It should also be noted that benchmark testing AI models, though helpful in understanding just how powerful they are, is imperfect. The datasets used to benchmark models have been found to be part of a model’s training, meaning the model already knows the answers to the questions evaluators will ask it.

Benchmark testing shows both sizes of Llama 3 outperforming similarly sized language models.

Meta says human evaluators also marked Llama 3 higher than other models, including OpenAI’s GPT-3.5. Meta says it created a new dataset for human evaluators to emulate real-world scenarios where Llama 3 might be used. This dataset included use cases like asking for advice, summarization, and creative writing. The company says the team that worked on the model did not have access to this new evaluation data, and it did not influence the model’s performance.

“This evaluation set contains 1,800 prompts that cover 12 key use cases: asking for advice, brainstorming, classification, closed question answering, coding, creative writing, extraction, inhabiting a character/persona, open question answering, reasoning, rewriting, and summarization,” Meta says in its blog post.

Llama 3 performed better than most models in human evaluations, says Meta.

Llama 3 is expected to get larger model sizes (which can understand longer strings of instructions and data) and be capable of more multimodal responses like, “Generate an image” or “Transcribe an audio file.” Meta says these larger versions, which are over 400B parameters and can ideally learn more complex patterns than the smaller versions of the model, are currently training, but initial performance testing shows these models can answer many of the questions posed by benchmarking.

Meta did not release a preview of these larger models, though, and did not compare them to other big models like GPT-4.