Good news, everyone! Meta has delighted AI enthusiasts with exciting release of the new LLM Llama version 3.1.

The new model is available in three sizes: 8B, 70B, and 405B parameters.

• 8B: A compact, extremely fast model suitable for deployment in any environment.

• 70B: A high-performance, budget-friendly model that supports numerous use cases.

• 405B: The premier foundational model, powering the most extensive range of applications.

Llama 3.1 Benchmarks

This is comparison of Llama 3.1 to its predecessor, a Llama 3 model.

Comparison of 3.1 with 3.0 shows significantly improved benchmarks, thanks to distilling the 405B model into smaller models.

Presenting a comparison of new models with competitors such as Gemini Gemma 2, Open AI GPT-4o, and Claude 3.5 Sonnet.

We now have a publicly available model at the level of GPT-4o that anyone can download!

We look forward to results from the Chatbot Arena and feedback from everyday users.

Get Access to Llama 3.1

New models are available for public access.

Method 1: Download the Models

To download the models, follow these simple steps:

1. Follow the link: https://llama.meta.com/llama-downloads/

2. Fill out the access request form with your information.

3. Select the models you want to access. You can choose from the new Llama 3.1, Llama Guard 3, Prompt Guard, as well as older models.

4. Accept Meta's Community License Agreement.

5. You will then see the download link for the models.

Please also read Llama 3.1 Download Instructions.

Llama 3.1 Models Available For Download

Meta has released a wide variety of new model variations, tailored to meet different user needs.

Pretrained:

• Meta-Llama-3.1-8B

• Meta-Llama-3.1-70B

• Meta-Llama-3.1-405B

• Meta-Llama-3.1-405B-MP16

• Meta-Llama-3.1-405B-FP8
  
  

Fine-tuned:

• Meta-Llama-3.1-8B-Instruct

• Meta-Llama-3.1-70B-Instruct

• Meta-Llama-3.1-405B-Instruct

• Meta-Llama-3.1-405B-Instruct-MP16

• Meta-Llama-3.1-405B-Instruct-FP8

• Llama-Guard-3-8B

• Llama-Guard-3-8B-INT8

• Llama-Guard-2-8B

• Llama-Guard-8B

• Prompt-Guard-86M

Some notes for 405B model:

Meta AI released several versions of the 405B model to accommodate its large size and provide various deployment options:

• MP16 (Model Parallel 16) - This is the full BF16 weights version. These weights can only be deployed across multiple nodes using pipelined parallel inference. A minimum of 2 nodes with 8 GPUs each is required for deployment.

• MP8 - This is also the full BF16 weights version but can be deployed on a single node with 8 GPUs using dynamic FP8 (Floating Point 8) quantization. Code for this deployment is available.

• FP8 (Floating Point 8) - This is the quantized version of the weights. These weights can be deployed on a single node with 8 GPUs using static FP8 quantization. Code for this deployment is also available.

• The 405B model requires approximately 750 GB and a minimum of two nodes (with 8 GPUs each) for inference in MP16.

Method 2: Use Llama 3.1 via ChatLabs

We always strive to add new models as quickly as possible, and this time is no exception. Llama 3.1 405b is available in ChatLabs!

To use the 405B model:

1. Go to ChatLabs website: https://labs.writingmate.ai

2. Sign up with your email.

3. Choose Llama 3.1 405b Instruct in the model list in the top right corner.

That's it! You can now use the newest LLM by Meta.

It will be helpful for you to know that ChatLabs also supports the comparison of various AI models in terms of cost, quality, performance, and response speed using the AI Split Screen mode feature. You can not only use the most advanced Llama 3.1 405B, but also compare it with any of the more than 30 other supported LLMs, including the most popular paid options.

Useful Links

• Official pressrelease: https://llama.meta.com

• Llama 3.1 Documentation: https://llama.meta.com/docs/overview

• Download the models from HuggingFace: https://huggingface.co/meta-llama

• Meta AI directory on GitHub: https://github.com/meta-llama

• GitHub Llama 3.1 Download Instructions: https://github.com/meta-llama/llama-models/blob/main/README.md

• Meta AI directory on Kaggle: https://www.kaggle.com/organizations/metaresearch/models