Startups Aren’t About Incremental Changes Says Lin Qiao As Fireworks AI Takes the Ten-Times Leap

Lin Qiao’s journey from deciphering ship blueprints as a child to leading one of the most dynamic AI infrastructure companies today is nothing short of extraordinary. As the CEO and co-founder of Fireworks AI, her philosophy—“A single model is not enough”—has not only shaped her company’s meteoric rise but has also redefined how businesses approach AI development.

In just six months, Fireworks AI has achieved what many companies take years to accomplish: processing over 150 billion tokens and generating more than one million images daily. This 100-fold increase in traffic underscores both the company’s technical innovation and the surging demand for AI solutions. But the story of Fireworks AI is as much about technology as it is about Qiao’s visionary leadership.

Growing up, Qiao spent hours with her father, a senior mechanical engineer, learning the intricacies of ship blueprints. “My dad built massive cargo ships from scratch, and I loved learning to interpret the intricate angles and measurements,” she recalls. That early exposure to engineering shaped her analytical mindset, but it was a high school programming assignment—a simple snake game—that ignited her passion for computer science.

Her academic journey took her to a Ph.D. in distributed systems and database management by 2005, followed by pivotal roles at IBM, LinkedIn, and later, Meta (formerly Facebook). At Meta, Qiao witnessed a seismic shift in AI infrastructure as companies transitioned from CPUs to GPUs. She led a team that grew from five to 300, creating transformative tools like Caffe2 and PyTorch.

“PyTorch isn’t just a framework,” Qiao explains. “It’s been foundational for robotics, self-driving cars, and even streaming services like Netflix for personalized content delivery.” The reach of PyTorch laid the groundwork for Qiao’s realization: businesses needed more accessible, scalable AI infrastructure.

Fireworks AI

In October 2022, just before ChatGPT’s launch, Qiao recognized an opportunity. “Companies wanted to prioritize AI but lacked the infrastructure, resources, and talent,” she observed. Inspired by PyTorch’s flame logo, she founded Fireworks AI with the vision of spreading that flame across industries.

Fireworks AI quickly gained momentum. With $25 million in seed funding from Benchmark and a recent $52 million Series B led by Sequoia, the company achieved a valuation of $552 million—a fourfold increase in just two years. Its user base has grown to over 23,000 developers, processing more than 60 billion tokens daily.

Their compound AI system, capable of handling diverse data formats like PDFs and images, and Whisper v3-large models—transcribing one hour of audio in just four seconds—are redefining efficiency. These systems achieve transcription speeds 900 times faster than real time, with a Word Error Rate of just 2.00% on the Librispeech Clean dataset.

The company’s Twitter post announcing this breakthrough sparked excitement, offering free trials for two weeks:
“We made Whisper 20x faster than OpenAI! Beta launching the fastest and most feature-complete audio APIs—transcribe ONE HOUR of audio in as little as 4 seconds. (900:1 transcription speed!).”

We made Whisper 20x faster than OpenAI*! Today, we’re beta launching the fastest and most feature-complete audio APIs – transcribe ONE HOUR of audio in as little as 4 seconds. (900:1 transcription speed!)

We’re offering it FREE for 2 weeks to celebrate launch – try it… pic.twitter.com/Q0PpolyJOZ

— Fireworks AI (@FireworksAI_HQ) December 9, 2024

Fireworks AI’s developer-first approach includes over 100 pre-trained models and the ability to fine-tune proprietary ones. With features like prompt caching and FireAttention, developers can integrate custom solutions with OpenAI-compatible APIs. Beyond cost efficiency—offering AI development 20 to 120 times cheaper than competitors—the platform ensures robust security through HIPAA and SOC2 compliance.

“Our customers first came to us for low-latency support,” Qiao says. “As their applications scaled, they needed solutions that didn’t break the bank.” Fireworks AI’s infrastructure meets these demands, making it indispensable for developers and businesses alike.

The Human Element

Behind the technical achievements is Qiao’s deeply human approach. As a mother of two teenagers, she is acutely aware of AI’s societal implications. “I worry about misleading or inappropriate content,” she admits. “Content safety is something the industry is just beginning to tackle.”

Her personal experiences have shaped her leadership style. “In the early stages of my career, I dealt with imposter syndrome. But over time, I built muscle for improvement, always thinking there’s room to grow,” she shares. This mindset drives her philosophy of bold innovation. “Startups aren’t about incremental changes. They’re ten-times leaps.”

Qiao’s hiring practices reflect this ethos. “I value aptitude over experience. I look for hunger, motivation, and a fast-learning mindset. Everything in this field is new, so the ability to adapt is critical.”

With the launch of DeepSeek V3 which is available on Fireworks Serverless and Enterprise, Fireworks AI continues to redefine what’s possible in the world of generative AI. DeepSeek V3, boasting 671B MoE parameters and 37B activated parameters, has quickly become a standout in both coding and reasoning. It has already earned top honors as the best-performing open model on platforms like Chatbot Arena and WebDev Arena, further solidifying its place at the forefront of AI innovation.

DeepSeek V3, a state-of-the-art open model, is now available on Fireworks Serverless and Enterprise!
🥇 SOTA open model for coding and reasoning
🥇 Best performing open model on Chatbot Arena and WebDev Arena
🧠 671B MoE parameters, 37B activated parameters
Congrats to the…

— Fireworks AI (@FireworksAI_HQ) December 31, 2024

Available for just $0.9 per million tokens, DeepSeek V3 brings unmatched performance with a 131K context size, blazing speeds of up to 30 tokens per second, and the promise of even faster optimization ahead. While the model is in its early release phase, Fireworks is already working hard to improve its speed and capabilities, with further evaluations and enhancements underway.

For Fireworks AI’s founder, Lin Qiao, this leap is just the beginning. As she notes, “Real-life communication goes beyond text. We need models that understand and generate images, audio, and other signals.” With DeepSeek now available through Fireworks, the company is one step closer to her vision of autonomous, self-organized generative AI applications—much like AlphaGo learning chess independently.