Fireworks AI
Fireworks AI is a generative AI platform built by the creators of PyTorch. It provides fast inference and training for open-source LLMs, vision models, and custom fine-tuned models through a developer-friendly API.
The platform offers three serving modes: serverless pay-per-token inference, on-demand GPU deployments, and reserved enterprise capacity. You can run hundreds of frontier models or deploy your own post-trained versions on a disaggregated inference engine optimized from custom kernels to memory management.
Fireworks also supports supervised fine-tuning, preference tuning (DPO), and reinforcement fine-tuning so teams can specialize open models on private data. Checkpoints deploy to production in seconds. The platform processes over 40 trillion tokens per day and is used by companies like Cursor, Vercel, Notion, Uber, and DoorDash.
Serverless inference with pay-per-token pricing and $1 in free credits to start
On-demand GPU deployments on H100, H200, B200, and B300 hardware
Fine-tune open models with LoRA or full-parameter SFT and DPO training
Reinforcement fine-tuning for specialized model post-training workflows
OpenAI and Anthropic compatible API so you can swap endpoints without rewriting code
Disaggregated inference engine delivering up to 4x higher throughput than general stacks
Model library with hundreds of text, vision, and embedding models ready to run
Enterprise deployments with reserved capacity, BYOC, and multi-region support
OpenAI-compatible API makes migration straightforward for existing integrations.
Disaggregated inference engine delivers up to 4x higher throughput than general-purpose stacks.
Fine-tuned models deploy at the same per-token price as base models.
Used in production by major AI companies including Cursor and Vercel.
Serverless pricing varies by model and requires checking the docs for current rates.
Enterprise pricing is not published and requires contacting the sales team.
Reinforcement fine-tuning costs scale with GPU hours, which can add up quickly.
Is Fireworks AI free to use?
Fireworks AI offers $1 in free credits when you sign up for serverless inference. After that, you pay per token for serverless usage or per GPU hour for on-demand and fine-tuning workloads.
What models can I run on Fireworks AI?
Fireworks AI provides access to hundreds of open-source text, vision, and embedding models. You can also fine-tune base models on your own data and deploy the resulting checkpoints at the same price as base models.
How does Fireworks AI pricing work?
Serverless inference is billed per token with postpaid billing. On-demand deployments are billed per GPU second on H100, H200, B200, or B300 hardware. Fine-tuning is priced per 1M training tokens for SFT and DPO, or per GPU hour for reinforcement fine-tuning.
Can I fine-tune models on Fireworks AI?
Yes. Fireworks AI supports LoRA and full-parameter supervised fine-tuning, preference tuning with DPO, and reinforcement fine-tuning. Fine-tuned models are served at the same price as their base models.
Is Fireworks AI compatible with OpenAI APIs?
Yes. Fireworks AI serverless inference is OpenAI and Anthropic compatible, so you can migrate existing integrations by changing the API endpoint URL.

