Fireworks AITranslation site

Fireworks AI Company Profile

Fireworks AI is a U.S. generative AI startup founded in 2022, focusing on providing large model fine-tuning, inference, and deployment services for enterprises and developers. Most of its founding team members come from Meta, Google and other big players, and its founder and CEO, Qiao Lin, graduated from Fudan University, has a PhD in computer science from the University of California, Santa Barbara, and has served as the head of Meta PyTorch, and also has technical work experience at LinkedIn and IBM. With the experience and technical strength accumulated in Meta for many years, the team is committed to helping enterprises rapidly realize AI transformation. fireworks AI developed a custom FireAttention inference engine, which reduces inference time by 12 times compared to the open source vLLM, reduces the cost of use, and has gained a lot of capital, with a valuation of 552 million dollars after the Series B round of financing in 2024 .

Core business and technology

• Large Model Service: Provides over 100 advanced text, image, audio, and multimodal macromodels covering large language models, image generation models, audio generation models, video generation models, embedded models, and more, dramatically optimized for latency, throughput, and cost.
• Model fine-tuning: Helps developers quickly customize models through ultra-fast LoRA fine-tuning technology, which takes only minutes from dataset preparation to querying the fine-tuned model, and the fine-tuned model can be seamlessly deployed into existing business processes.
• Reasoning Optimization: Enables semantic caching to avoid duplicate computation, specializes in capturing application workload patterns and building them into an inference stack that can automatically adapt to developer or enterprise workloads.
• Cost and efficiency: 12x reduction in inference time compared to traditional methods and 40x reduction compared to GPT-4; 140 billion tokens of data processed per day with 99.99% API uptime; RAG speed 9x higher than Groq, SDXL image generation 6x faster than other providers' averages, and presumably decoding speeds up to 1,000 tokens/sec; and Open Source Five times lower cost compared to the original model, and thirty times lower in the case of further fine-tuning.

Market Performance and Customers

• Financing: In July 2024, the company completed a $52 million Series B financing round led by Sequoia Capital, with participation from NVIDIA, AMD, MongoDB, etc., valuing the company at $552 million and totaling $77 million in financing.
• client base: Enterprises including Cresta, Cursor, Liner, DoorDash, Quora, and Upwork, as well as individual clients such as Cursor and Superhuman, customize quantitative solutions for specific use cases.

Team & Background

• Founder's backgroundQiao Lin, the founder and CEO, graduated from Fudan University, has a PhD in Computer Science from University of California Santa Barbara, was the head of Meta PyTorch, and has technical work experience at LinkedIn and IBM. Most of the team members are from Meta, Google and other big factories, with more than 1/3 Chinese members.
• technical accumulation: The team is committed to helping enterprises rapidly realize AI transformation with the experience and technical strength accumulated in Meta for many years.

Vision for the future

• Building a comprehensive knowledge access API: Work on creating a powerful API with precise calls to different models and APIs to enable access to full knowledge.
• Expanding Teams and Partnerships: Plans to use the new funding to expand the team and plans to expand partnerships with AI companies to drive the industry's shift to composite AI systems.