Fireworks.ai fine-tunes open-source AI models into APIs, simplifying integration for businesses, bridging the gap between LLMs and practical applications. (Pixabay)
A California-based startup, Fireworks.ai, is carving a niche in the artificial intelligence (AI) landscape with its innovative approach tailored for enterprises. Rather than constructing large language models (LLMs) or foundation models from the ground up, the company specializes in refining existing open-source models and transforming them into easily deployable Application Programming Interfaces (APIs). This method involves fine-tuning the models to narrow their focus on specific functionalities, thereby minimizing instances of AI misinterpretations and enhancing overall performance.
Founded by Lin Qiao, who also serves as the CEO, Fireworks.ai emerged from Qiao's extensive experience as Senior Director of Engineering at Meta, where she worked extensively with AI frameworks and platforms. Qiao and her team launched the startup in October 2022, aiming to streamline AI integration for businesses. In a recent discussion with TechCrunch, Qiao emphasized the company's core service of model fine-tuning. She explained, "Our service covers off-the-shelf open-source models, models we refine, or models clients can adjust themselves. All these options are accessible through our inference engine API."
Fireworks.ai stands out by acting as a facilitator between LLMs and business applications, offering a range of APIs designed for seamless integration. With an emphasis on API development, the company allows enterprise clients to easily incorporate any open-source AI model from its extensive library. Additionally, businesses can experiment with different models to find the best fit for their requirements.
Currently, Fireworks.ai boasts a repository of 89 open-source LLMs, including notable names such as Mixtral MoE 8x7B Instruct, Meta's Llama 2 70B Chat, Google's Gemma 7B Instruct, and Stability AI's Stable Diffusion XL. These models are available in two formats: serverless, which eliminates the need for hardware configuration or model deployment, and on-demand, tailored for dedicated deployments and served with reserved GPU configurations according to specific business needs.
For businesses opting for the on-demand format, Fireworks.ai offers three payment plans: Developer, Business, and Enterprise. The Developer plan operates on a pay-per-usage structure with a rate limit of 600 requests per minute, while the Enterprise tier provides customized pricing options with unlimited rate limits. The serverless format follows a per-token pricing model, with varying rates depending on whether the models are text-only, image-only, or multimodal.
