Fireworks ships inference, model-access and training products at Forge: a 30-turn agent task drops from 1 minute to 8 seconds

Fireworks released a set of products aimed at helping companies turn advances in open-weight models into intelligence built around their own expertise, data and tasks. Teams moving AI into everyday work faced agents that respond too slowly, models that need to perform better on company-specific tasks, and infrastructure that must meet enterprise requirements for cost, control and security. The company says the launches address those three practical barriers and span inference acceleration, model access, training and enterprise governance.

On the inference side, the target is the waiting time that accumulates across an agent's multiple model calls. An agent often makes several model calls to finish one task, and each call adds to the wait while running more agents raises capacity pressure and cost. Fireworks Ultra-fast, powered by NVIDIA Blackwell GPUs, is in public preview and demonstrated up to 1,500 tokens per second on DeepSeek V4.1 Flash; in a benchmarked 30-turn agent task, completion time fell from 1 minute to 8 seconds.

Cost-sensitive work is routed to a cheaper runtime, and a specialized model saves on token usage. Not every task needs immediate completion, yet data generation, evaluations and overnight research previously drew on the same high-cost inference capacity. Background Mode offers a lower-cost runtime at half price; multi-day Prompt Caching lowers costs for repeated context; Ember-1 used about 40% fewer tokens while maintaining comparable quality to Kimi K3 in Fireworks' evaluations.

Training and enterprise controls carry specialized models from experiment into production. One of the biggest barriers to training your own model is that writing evaluations could take weeks of data labeling and grading. No Eval is in private preview and turns production traces and product signals into task-specific evaluations; Specialized Model Inference is in public preview, serving LoRA adapters serverlessly starting with GLM 5.3 Flash; Fireworks Sovereign is in private preview, bringing the inference engine to customers' cloud, neocloud or data centers, and the company says it already manages over 25,000 GPUs in customer environments.

The report covers the products Fireworks announced at Forge and the company's own demonstration figures, including Ultra-fast speed on DeepSeek V4.1 Flash, the 30-turn agent task completion time, Background Mode at half price, the Ember-1 token comparison, and each product's preview or general-availability status.

The speed and cost figures come from the company's own demonstrations and evaluations; the 1 minute to 8 seconds on a 30-turn task, 1,500 tokens per second and Ember-1's roughly 40% fewer tokens are not accompanied by test conditions or independent validation, and most capabilities are in public or private preview with limited general availability.