Fireworks AI
Verified May 27, 2026Fireworks AI focuses on serverless inference for open-weights models with a strong emphasis on throughput. The catalog skews toward production-grade Llama, Qwen, and Mixtral variants, with first-class support for function calling, JSON-constrained output, and vision endpoints.
Pricing is per-token for serverless and per-GPU-second for on-demand deployments, which makes capacity planning straightforward for spiky workloads. Fireworks is one of the few providers that publishes both speculative-decoding throughput numbers and median time-to-first-token on each model page, so the published specs match observed performance more often than competitors.
Inference runs from US regions with HIPAA-eligible deployments on enterprise plans; EU residency requires a dedicated cluster rather than the shared serverless tier.
Calculate cost for your workload
Plug in your monthly tokens — get the actual bill on every provider serving each model.
Open calculator