Hosting and adapting open models
Fireworks AI provides inference and training services for open models. Developers can query models through a serverless API or use dedicated deployments, with an OpenAI-compatible interface for text inference. The platform also offers tools for training specialized variants and routing workloads among models. [1][3]
This combines two roles: operating models supplied by others, and helping teams adapt models for particular tasks. A common request interface reduces integration work, while training changes the model itself. The interface and deployment documentation establish those product choices, rather than guaranteeing equivalent behavior or economics across models. [1][3]
Ember-1 added an in-house model
On September 23, 2026, Fireworks released Ember-1 as a research preview. It post-trained Kimi K3 to shorten reasoning traces for coding and agent workloads, using its own training platform. That made Fireworks the developer of this specialized variant, in addition to its role as a serving provider. [2]
The launch specified an initial two-week serverless research-preview window, with permanence depending on demand. Fireworks' token-saving figures came from its own benchmarks and customer tests; they are not a fixed reduction on every task. The historical preview announcement alone does not establish ongoing API availability today. [2]