Fireworks AI
Fireworks AI is an inference and training platform for open models, offering fast serverless APIs, fine-tuning and dedicated GPU deployments.
About this tool
Fireworks AI runs open-source models such as DeepSeek, Qwen, Kimi and OpenAI’s open-weight models behind fast APIs. Developers can call serverless endpoints and pay per token for input, output and cached tokens, with no setup and no cold starts. Fireworks positions it as a substitute for proprietary model APIs that can send each request to whichever open or proprietary model suits it.
The platform also covers customisation. Teams can fine-tune open models with supervised or preference methods, either through a guided flow that proposes a plan and cost for approval or by supplying their own configuration and training logic. Fine-tuned models are served at base-model prices, and dedicated GPU deployments billed per second offer higher rate limits and lower latency at scale.
What you can do
Serverless inference
Call popular open models through per-token APIs without managing infrastructure or cold starts.
Fine-tuning
Customise open models with SFT or DPO, priced per million training tokens.
Dedicated deployments
Run models on reserved GPUs billed per second for higher throughput and rate limits.
Model routing
Send each request to the open or proprietary model best suited to it, including coding workloads.
Who is it for?
- A coding-assistant startup moves part of its traffic from a closed API to an open model to cut costs.
- A support platform fine-tunes an open model on its ticket history.
- An enterprise deploys a large open model on dedicated GPUs for predictable latency.
Before you choose
Billing is prepaid with usage deducted from a balance; set auto-reload or monthly limits to avoid service interruption or overspending.
Vendor savings estimates depend on workload; benchmark quality and cost on your own tasks before switching models.
Frequently asked questions
How does Fireworks AI charge?
Serverless use is billed per token from a prepaid balance, fine-tuning per million training tokens, and dedicated GPUs per second of use.
Which models can I use on Fireworks AI?
Fireworks hosts many open models, including large ones such as DeepSeek, Qwen and Kimi models and OpenAI’s open-weight models.
Do fine-tuned models cost more to serve?
No. Fireworks states that fine-tuned models are served at the same price as their base models on its serverless platform.
Check the official website for current features, pricing, and terms.
Similar tools
Together AI
CodingTogether AI is a cloud platform for running, fine-tuning and serving open-source AI models, from serverless APIs to dedicated endpoints and GPU clusters.
Replicate
CodingReplicate runs open-source and proprietary AI models through a cloud API, and lets developers fine-tune models or deploy their own without managing GPUs.
Modal
CodingModal is a serverless cloud for AI workloads: run inference, batch jobs, fine-tuning, sandboxes and GPU notebooks from Python code, billed per second.
Groq
CodingGroq is an inference cloud built around its own LPU chips, offering fast, OpenAI-compatible API access to open AI models for developers and enterprises.
DeepSeek
CodingDeepSeek develops large language models, including open-weight reasoning models, offered through a chat app and an OpenAI-compatible developer API.
Hedra
VideoHedra builds visual AI models and infrastructure, offering managed inference for 100+ models, fine-tuning and GPU sandboxes alongside its creative app.