FIREWORKS AI

Lin Qiao spent years at Meta running PyTorch, the open-source framework that most of the AI world uses to build models. She had a front-row seat to a coming problem.

Building a model is one thing. Running it fast and cheap for millions of users is another, and it is brutally hard.

In late 2022 she left with fellow PyTorch engineers Dmytro Dzhulgakov and Dmytro Ivchenko to fix exactly that. They founded Fireworks AI in Redwood City.

The pitch was simple. Let companies plug into open-source AI models and run them quickly without building their own infrastructure.

Fireworks AI runs other people's AI models for them, fast. Think of it as a power grid for AI.

A company that wants to add AI to its app does not want to buy GPUs, tune servers, and babysit models. Fireworks does all of that and charges based on usage, per token processed.

It gives customers access to hundreds of open-source models across text, images, and audio, and makes them run faster and cheaper than doing it yourself. The specialty is inference, which is just the running of a trained model to actually answer a request.

By 2025 it was handling more than 10 trillion tokens every single day.

The core product is the Fireworks inference platform, a service where developers can run hundreds of open-source AI models through a simple interface without managing any hardware. It handles text, image, and audio models, and offers fine-tuning so companies can customize a model on their own data.

It is built for production, meaning real apps serving real users at scale, not just experiments. The company also pushes tools for building compound AI systems, where several models and functions work together to answer a single request.

Fireworks grew by being the fast, neutral option in a market obsessed with speed and cost. Instead of building its own flagship model to fight OpenAI, it stayed the layer underneath, tuning open-source models so they run cheaper.

That made it a friend to everyone rather than a rival. It leaned on its founders' PyTorch reputation to win over developers who trust that team on performance.

It also pushed the idea of compound AI systems, chaining several models and tools together to do a job. Big names like Uber, Shopify, and DoorDash became customers, and word spread that Fireworks was the quiet workhorse behind a lot of AI features.

Fireworks lives in a brutal, crowded market. It competes with Together AI, Baseten, and the big cloud providers like Amazon and Google, all fighting to be the place you run AI models.

The work is also a race to the bottom on price, because inference is becoming a commodity and margins are thin. On top of that, its whole business depends on open-source models staying good enough to compete with the closed ones from OpenAI and Anthropic.

If the best models all go private and locked down, the neutral inference layer gets squeezed. Staying the fastest and cheapest is a treadmill that never stops.

Fireworks raised money from a who's who of tech. The $25 million Series A was led by Benchmark, with partner Eric Vishria joining the board, plus Sequoia, Databricks Ventures, and angels like former Snowflake CEO Frank Slootman and Scale AI's Alexandr Wang.

In July 2024 the $52 million Series B was led by Sequoia at a $552 million valuation, with NVIDIA, AMD, and MongoDB joining. Then in October 2025 came the big one.

A $250 million Series C led by Lightspeed Venture Partners with Index Ventures and Evantic, pushing the valuation to $4 billion and total funding to around $327 million.