Today's Open Source (2024-07-23): Llama-3.1-405B Magnet Link Available for Download

Explore today's AI open-source highlights: Meta's Llama 3.1, NVIDIA's Minitron and Audio Flamingo, Aide for VSCode, MambaInLlama, and Prelude for LLM prompts.

Here are some interesting AI open-source models and frameworks I wanted to share today:

Project: Llama 3.1

Llama 3.1 is Meta's upcoming multilingual large language model with 8B, 70B, and 405B parameter versions.

It has a context of 128k and supports eight languages, including English, French, and German, but not Chinese.

Llama 3.1 performs excellently on various industry benchmarks, and uses advanced Transformer architecture, and autoregressive language modeling techniques.

The 405B parameter version is available via a magnet link:

https://pastebin.com/9jGkYbXY

Project: Minitron

Minitron is a series of small language models (SLMs) open-sourced by NVIDIA, pruned from the Nemotron-4 15B model.

The model's embedding size, attention heads, and MLP dimensions were pruned, followed by training through distillation, resulting in Minitron 8B and 4B models.

This method requires up to 40 times less training data than training from scratch, saving 1.8 times the computational cost.

Minitron models achieve a 16% higher MMLU score than those trained from scratch and perform comparably to other community models like Mistral 7B, Gemma 7B, and Llama-3 8B, outperforming some earlier compression techniques.

https://huggingface.co/nvidia/Minitron-4B-Base

https://huggingface.co/nvidia/Minitron-8B-Base

Project: Audio Flamingo

Audio Flamingo is an open-sourced audio language model by NVIDIA that features strong audio understanding, few-shot learning, and multi-turn dialogue capabilities.

The project uses advanced training techniques, architecture design, and data strategies to enhance these capabilities and set new benchmarks in various audio understanding tasks.

https://github.com/NVIDIA/audio-flamingo

https://arxiv.org/abs/2402.01831

https://huggingface.co/nvidia/audio-flamingo

Project: Aide

Aide is a VSCode extension that helps developers master any code with one-click annotations and language conversion.

It offers code conversion, detailed annotations, batch copying, and variable renaming, and supports multi-file operations and custom commands, significantly improving code readability and development efficiency.

https://github.com/nicepkg/aide

Project: MambaInLlama

MambaInLlama is a hybrid model tool for distilling and accelerating large Transformer models, aiming to reduce resource consumption and improve efficiency while maintaining generation quality. It is suitable for optimizing both foundational and chat models.

https://github.com/jxiw/MambaInLlama

Project: Prelude

Prelude is a development tool for creating long-context window LLM prompts from code repositories.

It's useful for enhancing code spread across multiple files and directories with LLMs.

Prelude generates prompts containing the file tree and the content of specified directories, automatically copying them to the clipboard or saving them to a file.

https://github.com/aerugo/prelude