Fireworks Ultrafast is a public preview high-speed inference mode delivering over 1,500 tokens per second for latency-sensitive interactive applications, using NVIDIA Blackwell, along with multi-day prompt caching, background mode for long-running agents, and higher rate limits.
Fireworks Ultra-fast is an inference mode for time-sensitive workloads, including agents that make multiple model calls. Fireworks says it is powered by NVIDIA Blackwell GPUs and, in its public preview, reached up to 1,500 tokens per second on DeepSeek V4.1 Flash. The company reports that a benchmarked 30-turn agent task took eight seconds with Ultra-fast, compared with one minute otherwise. Fireworks also describes multi-day prompt caching and higher rate limits as related ways to support persistent or concurrent agent workloads. The speed and benchmark figures are Fireworks’ own claims.
Be the first Cosign
Add it to your Products I use list. You can also share why you use it.
