Fireworks AI is a generative AI inference and training cloud founded in 2022 by veterans of Meta's PyTorch team, led by CEO Lin Qiao, former head of PyTorch at Meta, with co-founders including Dmytro Dzhulgakov and Benny Chen. The company is headquartered in Redwood City, California, and positions itself as offering frontier performance with open model economics; it reports processing more than 30 trillion tokens per day.
For inference, Fireworks offers serverless pay-per-token access with Priority and Fast tiers, on-demand dedicated GPU deployments across regions, and reserved capacity, all through OpenAI- and Anthropic-compatible APIs. Its model library includes more than 30 open and proprietary models such as DeepSeek, GLM, Qwen3, Kimi K2, Gemma, and FLUX.1 image models.
For fine-tuning, the platform supports supervised fine-tuning and DPO in both LoRA and full-parameter modes, multi-LoRA serving that deploys many fine-tuned adapters cheaply on one base model, and reinforcement fine-tuning pipelines. Fine-tuning is priced per million training tokens, from $0.50 for LoRA SFT on models up to 16B parameters, scaling with model size and method. On-demand GPUs are billed hourly (H100 at $7/hour, B200 at $10/hour).
New accounts receive $1 in free credits; enterprise plans are available through sales. Typical uses include production LLM applications, agents, code assistants, and image generation at scale.