Fireworks unveiled new products at its Forge conference. Ultra-fast inference on NVIDIA Blackwell GPUs reaches 1,500 tokens per second. Background Mode cuts costs by half for non-urgent tasks. Multi-day Prompt Caching lowers costs for repeated context. Ember-1 uses 40% fewer tokens at roughly a fifth of Opus 5’s cost. Fireworks Nexus brings open models into existing tools like Codex and Claude Desktop. Nexus Desktop and CLI route models without terminal use. Nexus Pro starts at $8 per month and Teams at $10 per seat. Elastic Burst and higher serverless rate limits are now generally available. No Eval in private preview turns production traces into task-specific evaluations. Specialized Model Inference serves LoRA adapters via serverless API. Free training on Muse Glimmer 30B runs through November 30. Model governance, spending controls and audit logs are now generally available.
What changed
Agents ran at uniform speed and cost regardless of workload urgency or repetition.
What it unlocks
Running faster agents for time-sensitive tasks and cheaper agents for background work in the same stack.
- 1,500 tokens per second on Ultra-fast
- 30-turn agent task: 1 minute to 8 seconds
- Background Mode at half price
- Ember-1: 40% fewer tokens, ~1/5 cost of Opus 5
What you need to act on it
- public preview for Ultra-fast and Specialized Model Inference
- private preview for No Eval and Fireworks Sovereign
- Nexus Desktop requires Codex, Claude Desktop or OpenCode
Sources