Replicate Intelligence #8: Meta Releases Llama 3.1 405B, Mistral Unveils Large 2, and Meta Open-Sources AI Agent Toolkit

Replicate published its weekly bulletin, Replicate Intelligence #8, on July 26, 2024, highlighting major open-source AI releases and platform updates.

Llama 3.1

Meta released the Llama 3.1 generation, including a 405 billion parameter model alongside updated 8B and 70B versions. According to the source, the 405B model features a 128,000-token context window, multilingual support, and built-in tool use capabilities. Meta's research details extensive use of synthetic data, novel fine-tuning approaches, and techniques for handling long contexts. The updated Llama license permits synthetic data creation for training other AI models with certain restrictions.

Mistral Large 2

Mistral AI unveiled Mistral Large 2, a 123 billion parameter model released under a Research License. The model features a 128,000-token context window and reportedly matches Llama 3.1 405B performance on some tasks, with particular strength in coding and mathematical reasoning.

Tools and Safety

Meta open-sourced a toolkit for building AI agents with Llama 3.1, designed to break down complex tasks and use built-in and custom tools with configurable safety via Llama Guard. Meta also released PromptGuard, a classification model based on mDeBERTa-v3-base that detects malicious prompts by classifying inputs as BENIGN, INJECTION, or JAILBREAK.

Replicate Platform Update

Replicate added a new API endpoint enabling programmatic search of all public models on the platform using plaintext queries with paginated JSON responses.

Source: Replicate Blog