Provider pricing

Fireworks AI AI API pricing

Fireworks AI runs open-weight models on a custom inference stack (FireAttention, speculative decoding) that delivers some of the lowest latencies in the industry. The catalog covers Llama 4, DeepSeek V4, Qwen3, Mixtral, and Nemotron, plus image, audio, and embedding models — all behind an OpenAI-compatible API.

Founded2022
HeadquartersRedwood City, CA
Tracked models8

Best for

Latency-sensitive production workloads on open-weight models — agents, voice, real-time chat.

Strengths

  • Industry-leading latency via FireAttention + speculative decoding
  • Function calling, JSON mode, and structured output on open-weight models
  • On-demand fine-tuning + LoRA serving at no extra inference cost
  • Compound AI: chain models, retrieval, and tools server-side

Tracked Fireworks AI models

ModelInput $/MOutput $/MContextTierBest for
Llama 4 Scout (Fireworks AI)$0.090$0.2501,000,000BudgetFireworks-hosted Llama 4 Scout — ultra-fast inference
Llama 4 Maverick (Fireworks AI)$0.180$0.5001,000,000Mid-tierFireworks-hosted Llama 4 Maverick — low-latency reasoning
Qwen3 32B (Fireworks AI)$0.250$0.700128,000Mid-tierFireworks-hosted Qwen3 32B optimized for throughput
Mixtral 8x22B (Fireworks AI)$1.80$5.4064,000FlagshipFireworks-hosted Mixtral 8x22B with speculative decoding
Mistral Small 3.1 (Fireworks AI)$0.130$0.500128,000BudgetFireworks-hosted Mistral Small 3.1 for high-volume agents
Llama 3.3 70B (Fireworks AI)$0.900$0.900128,000Mid-tierFireworks-hosted Llama 3.3 70B — low-latency inference
DeepSeek R1 (Fireworks AI)$3.00$8.00128,000ReasoningFireworks-hosted DeepSeek R1 with optimized throughput
Qwen 2.5 Coder 32B (Fireworks AI)$0.900$0.900131,072Mid-tierFireworks-hosted Qwen 2.5 Coder 32B for agentic dev tools