Fireworks.ai is a California-based synthetic intelligence (AI) startup that’s providing a novel resolution for enterprises. The AI agency doesn’t construct massive language fashions (LLMs) or basis fashions from scratch however fine-tunes open-source fashions and converts them into an Software Programming Interface (API) to assist companies deploy the AI capabilities in a seamless trend. The fine-tuning reduces the scope of the AI mannequin and focuses it on a selected performance. This permits them to scale back cases of AI hallucinations and enhance the capabilities of the mannequin considerably.
The AI agency was co-founded by Lin Qiao who additionally holds the seat of the CEO within the firm. After serving because the Senior Director of Engineering at Meta and dealing with AI frameworks and platforms, Qiao and her group based the startup in October 2022, as per her LinkedIn profile. In a dialog with TechCrunch, she defined the enterprise mannequin of Fireworks.ai, highlighting the fine-tuning service they supply. She stated, “It may be both off the shelf, open supply fashions or the fashions we tune or the fashions our buyer can tune by themselves. All three varieties might be served by means of our inference engine API.”
This places the agency in a novel place the place whereas it’s not innovating on the basis mannequin degree, it’s bridging the hole between an LLM and a business-ready product that may be deployed seamlessly. With a major deal with constructing APIs, Fireworks.ai lets its enterprise shoppers plug and play any open-source AI mannequin in its catalogue. As per the report, the corporate additionally lets companies experiment with totally different AI fashions to decide on the one that matches their wants.
At current, the startup claims to comprise 89 open-source LLMs corresponding to Mixtral MoE 8x7B Instruct, Meta’s Llama 2 70B Chat, Google’s Gemma 7B Instruct, Stability AI’s Steady Diffusion XL, and extra. The AI agency presents the fashions in both serverless format that doesn’t require companies to configure {hardware} or deploy fashions, or as on-demand fashions which can be found for devoted deployments, served on reserved GPU configurations in line with enterprise wants.
For the on-demand format, Fireworks.ai has three fee plans — Developer, Enterprise, and Enterprise — the place the Developer plan comes with a pay-per-usage construction and a price restrict of 600 requests per minute, the Enterprise tier has customized pricing presents and limitless price limits. The serverless format is billed at a per-token pricing plan the place totally different fashions, relying on whether or not they’re text-only, image-only, or multimodal, will fetch a distinct worth.
