April 2024 Summaries

Fireworks has announced the integration of Meta's open-source Llama 3 models, which are available with 8 billion (8B) and 70 billion (70B) parameters, into their platform, highlighting improved speeds for inference and fine-tuning capabilities. These models, which excel in language processing benchmarks, are part of a broader initiative by Meta to promote open and responsible AI development, with plans to release even larger models in the future. Fireworks offers businesses access to these models through a serverless inference stack, enabling cost-effective deployment of numerous LoRA adapters. Companies like Quora and Sourcegraph are already utilizing Fireworks for their AI-driven applications. By providing a seamless setup process through their platform, Fireworks aims to make state-of-the-art AI accessible to developers and enterprises, supporting both dedicated and shared infrastructure for enhanced performance.

Fireworks AI has partnered with Stability AI to offer the advanced image generation models Stable Diffusion 3 (SD3) and Stable Diffusion Turbo (SD3-turbo) through their distributed inference service. These models, which improve on previous versions, leverage latent diffusion to generate high-quality images from text descriptions, surpassing competitors like DALL·E 3 and Midjourney v6 in prompt adherence and typography. By using the new Multimodal Diffusion Transformer (MMDiT) architecture, SD3 enhances text understanding and spelling capabilities. Fireworks AI has benchmarked SD3 to achieve impressive speeds, with SD3-turbo further optimizing efficiency. The blog provides a tutorial on accessing the SD3 API using Python, guiding users through setting up their environment, authenticating, and generating images with customizable parameters. Fireworks AI aims to democratize AI for developers and businesses by providing fast and reliable image and text generation services.