A keynote packed with exciting announcements by @FireworksAI_HQ during its Forge event: - Fireworks Sovereign: The Fireworks stack in any environment you need it - Ember Next-gen: Back in September, Fireworks Research released Ember-1, a specialized model that delivers Kimi K3's quality with 40% fewer tokens. Built on Kimi K3, it learned to cut unnecessary reasoning while keeping the thinking that matters. Today, Fireworks announces its next-gen: based on US open models, it targets 30x more efficiency than Ember-1 - No-GPU: Pay-per-token training and inference. No need to worry about the underlying infrastructure - No-evals: Evals that build themselves from traces and product signals - Ultra-loop: Recursive self-improvement for inference. Every experiment feeds a shared knowledge base, so each loop starts smarter than the last - Ultra-fast: Models can reach 1,500–2,000 tok/s - Ultra-efficient: Persistent agents (e.g., personal assistants) and background agents (e.g., swarms) are the emerging use cases in the industry. To support those: - Multi-day prompt caching: Builds on the already industry-leading cache hit ratio (95%+) - Background-mode: 50%+ lower cost for agents that can afford to wait Incredible work, congratulations to the team! Source: Forge 2026