Fireworks Inference
Serving for frontier open models and for a customer's own post-trained versions of them, on an inference engine tuned at each layer.
Find alternatives to Fireworks Inference
What it does
- Model inference
- Model hosting
Features 5
- On-DemandDedicated deployments for post-trained models with multi-region support and custom performance optimizations.
- Reserved CapacityGuaranteed capacity, higher quotas and early access to the newest regions and hardware, with BYOC.
- Disaggregated prefill and decodeSplits prefill and decode onto separate pools so each scales independently.
- KV cache and routingDisaggregated KV caching and prompt-aware routing keep long prompts and multi-turn sessions fast.
- Multi-node expert parallelismServes frontier MoE models across nodes with composable parallelism.
Sources
- Pricing
-
Fireworks publishes per-token and per-GPU-hour rates; each surface is metered separately.
- Description
-
Official documentation · 6 Sep 2026
Summarised from the page's own title, <h1> and meta description, fetched 2026-09-06.
- Constraints
-
Official documentation · 2 Oct 2026
Derived from the security facts on this page, from the same source.
- Security
-
Official documentation · 2 Oct 2026
Platform-wide: Fireworks docs Data Security page plus Custom SSO and Audit Logs pages; covers inference and training products, not Nexus or speech recognition (not named). HIPAA described as 'compliant' only, no BAA stated so omitted. ZDR is the vendor's own policy for open models; SSO SAML and audit logs are Enterprise features. Training is from the privacy notice: 'We do not use your prompts, training data, or API inputs to train or improve our AI models without your explicit opt-in.' trust.fireworks.ai renders empty to curl.
- Sold within
-
Fireworks publishes per-token and per-GPU-hour rates; each surface is metered separately.
- URL
-
Official documentation · 6 Sep 2026
Fetched 2026-09-06: HTTP 200, <title> 'Inference | Fireworks', <h1> 'The highest performance inference for your specialized intelligence.'.
- Interfaces
-
Official documentation · 12 Oct 2026
Page: 'OpenAI and Anthropic compatible, so you can migrate by swapping a URL'.
- Modalities
-
Official documentation · 12 Oct 2026
Page: 'Embeddings models take text as input and output a vector of floating point numbers'.
- Intended for
-
'Built for any scale: From solo developers to Cursor-scale workloads' (developer); 'Reserved Capacity: guaranteed capacity...supports increased deployment flexibility with BYOC' (enterprise)
Also from Fireworks AI
-
Fireworks AI Model API
Hosted API for serving and fine-tuning open-weight models, billed per token.
-
Fireworks Nexus Model API
Routing layer for coding harnesses such as Claude Code, Codex and OpenCode that scores each request and sends it to an open model or your closed provider to cut cost.
-
Fireworks Training API service
Training and retraining of custom models on Fireworks, offered across several training surfaces and served on the same platform.
-
Fireworks Speech Recognition API service
Real-time and batch speech-to-text on Fireworks, aimed at voice workflows that need low-latency transcription at scale.
-
FireOptimizer Infrastructure service
FireOptimizer optimizes model inference and supports fine-tuning with users' data.
-
Fireworks Sovereign Infrastructure service
Managed inference on GPUs the customer owns: Fireworks qualifies the cluster, deploys the models and is responsible for upgrades, monitoring and incident response.
Open Fireworks AI in the directory
Something wrong here? Send a correction — quote this product id: fireworks-inference.