Fireworks AI
Train, customize, and serve open generative AI models with serverless, dedicated, and reserved deployment options.
About Fireworks AI
Fireworks AI is an infrastructure platform for teams that train, customize, and serve open generative AI models. Its model library offers text, vision, embedding, and reranking options, while the inference service supports serverless access, dedicated on-demand deployments, and reserved capacity. Serverless inference uses postpaid per-token billing, with Fast and Priority tiers referenced on the pricing page; cached input and batch inference are generally priced at half of standard serverless rates. Teams can also fine-tune supported models or move into configuration-led and custom training workflows as their needs mature. The service is most relevant to developers who need API access to multiple open models, want to deploy a trained checkpoint quickly, or need dedicated capacity for higher limits. Adoption still requires checking each model listing for current price, context length, and supported functions because availability differs by model. Costs can also span tokens, GPU time, or negotiated reserved capacity, so workload shape matters when comparing deployment modes.
Best for
Best use cases
Prototype with serverless inference before choosing dedicated capacity
Customize a supported open model with private training data
Run latency- or throughput-sensitive workloads on on-demand deployments
Compare before choosing
Compare model-specific token prices and supported functions before selecting a serverless model
Compare per-token serverless billing with per-GPU-second on-demand billing for the expected workload
Confirm whether the required model supports serverless inference, fine-tuning, function calling, or image input
Features
Serve open models through an API
Use the model library to access supported text, vision, embedding, and reranking models without setting up inference infrastructure.
Fine-tune supported models
Customize eligible open models with your data and deploy the resulting checkpoint through the same platform.
Run dedicated inference
Move workloads to on-demand GPU deployments or reserved capacity when dedicated resources and higher limits are required.
Pros
- Serverless access removes inference setup and avoids cold starts, which shortens the path from model selection to API testing.
- The same platform covers fine-tuning and deployment, reducing handoffs between model customization and production serving.
- Serverless, on-demand, and reserved options let teams align infrastructure with early experiments, dedicated workloads, or guaranteed capacity.
Cons
- Prices and supported functions vary by model, so teams must review individual listings before standardizing a workflow.
- Usage may be billed by tokens or GPU seconds depending on deployment mode, making cost comparisons workload-specific.
- Some models do not support serverless inference, fine-tuning, image input, embeddings, or reranking, which can limit model portability.
Pricing Plans
Serverless Inference
- Zero setup and no cold starts
- Fast and Priority options
- Cached input and batch inference generally priced at 50% of standard serverless rates
Embeddings up to 150M parameters
- Embedding inference
- Rate applies to base models up to 150M parameters
- Postpaid usage billing
Embeddings 150M–350M parameters
- Embedding inference
- Rate applies to base models from 150M to 350M parameters
- Postpaid usage billing
* Pricing information is for reference only. Please refer to the latest announcement on Fireworks AI's official website.