Meta releases AI model “Llama 3.3”, 70B model exhibits performance comparable to Llama 3.1’s 405B model – GIGAZINE

Meta is a large-scale language modelLlama 3.3” was released on Saturday, December 6, 2024. At the time of article creation, a model with 70B parameters has been released, and Amazon’s basic model “Amazon Nova ProIt is being promoted as having lower cost and higher performance than the previous generation model, and that it has performance comparable to the previous generation’s higher-end models.

Llama
https://www.llama.com/

llama-models/models/llama3_3/MODEL_CARD.md at main · meta-llama/llama-models · GitHub
https://github.com/meta-llama/llama-models/blob/main/models/llama3_3/MODEL_CARD.md

Llama 3.3 70B is trained on over 15 trillion tokens collected until December 2023. The context length is 128,000 and the supported languages ​​are English, German, French, Italian, Portuguese, Hindi, Spanish, and Thai.

Below is a table listing the benchmark results of Llama 3.3 70B and various AI models. Llama 3.3 70B outperforms Llama 3.1 70B in most tests, and also outperforms Amazon Nova Pro in MMLU, which evaluates overall performance, and MATH, which evaluates math-related performance. . In addition, in some benchmarks such as coding performance and math performance, it has exceeded the previous generation top model Llama 3.1 405B.

Meta claims that Llama 3.3 70B uses cost-effective inference to achieve performance similar to Llama 3.1 405B while being able to run locally on a typical developer workstation.

Llama 3.3 70BMeta’s official websiteIt is available at , and also distributed by Hugging Face. Please note that the license isLlama 3.3 Community License” and the number of monthly active users is less than 700 million, commercial use is possible.

meta-llama/Llama-3.3-70B-Instruct · Hugging Face
https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct