Fireworks unveils inference at up to 1,500 tokens per second and deployment on customer infrastructure
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
Companies gain more ways to accelerate agents, reduce their cost, and run open models inside their own perimeter.
At its Forge conference, the company opened a public preview of Ultra-fast on NVIDIA Blackwell: according to Fireworks' measurements, DeepSeek V4.1 Flash produces up to 1,500 tokens per second, and a 30-step test agent task fell from one minute to eight seconds. Background Mode runs eligible delayed tasks at half price, while multi-day caching reduces repeated context processing. Nexus Desktop brings a choice of open and closed models to Codex, Claude Desktop, and OpenCode; Specialized Model Inference serves LoRA adapters through serverless inference. Fireworks Sovereign is in private preview and moves inference and training into a customer's cloud or data center under its security policies.
Sources 2