Fireworks Unveils Ultra-fast Inference, Nexus, and Sovereign at Forge

Details

Fireworks Ultra-fast, powered by NVIDIA Blackwell GPUs, is in public preview and delivers speeds up to 1,500 tokens per second on DeepSeek V4.1 Flash, reducing a 30-turn agent task completion time from 1 minute to 8 seconds.
Fireworks Nexus brings open and closed models into existing tools like Codex and Claude Desktop, with Nexus Desktop launching alongside the CLI; Pro plans start at $8/month and Teams at $10/seat/month.
Fireworks Sovereign, now in private preview, runs the inference engine inside customer clouds or data centers, managing over 25,000 GPUs and offering technology that can double tokens per GPU dollar.
New training and evaluation tools include No Eval (private preview) for generating task-specific evaluations from production traces, and Specialized Model Inference (public preview) for serving LoRA adapters on GLM 5.3 Flash with per-token pricing.
Background Mode offers a lower-cost runtime at half price for non-urgent workloads, while Ember-1 uses about 40% fewer tokens than Kimi K3 with comparable quality.

Key point