Fireworks Forge Adds Nexus Desktop and New Inference Previews

Fast interactive work and overnight processing put different demands on an inference service.

Fireworks announced a range of Forge products on November 3, including Nexus Desktop and new options for serving specialized models. The announcements cover several release stages, so teams need to distinguish a launched service from a public or private preview.

Nexus Desktop connects existing tools

Nexus Desktop connects supported applications including Codex, Claude Desktop, and OpenCode to Nexus model routing. Fireworks lists Pro plans starting at $8 per month and Teams plans starting at $10 per seat per month. Those starting subscription prices should not be read as a promise that every inference workload is included without further charges.

The company also describes Background Mode as a half-price option for eligible work that can tolerate delays, such as evaluation-data generation. A team comparing that route with interactive inference needs to include its completion deadline in the cost comparison.

Check the release stage of each capability

Capability	Availability described at Forge
Ultra-fast inference	Public preview
Specialized Model Inference	Public preview; serverless LoRA adapters start with GLM 5.3 Flash
No Eval	Private preview
Fireworks Sovereign	Private preview
Model governance, spending controls, exportable audit logs	Generally available
Elastic Burst	Generally available when spare capacity exists

Specialized Model Inference serves trained LoRA adapters through a serverless API with per-token pricing. Fireworks says idle models don’t incur a serving charge. No Eval uses captured production runs and product feedback to help teams build evaluations for their particular tasks.

Choose the deployment constraint first

Sovereign is aimed at organizations that want inference and training inside infrastructure they control. Its private-preview status means you should treat it as an access discussion rather than a universally available deployment option.

For an initial evaluation, choose a single workload and fix its required output quality. Record the model and configuration, total cost, and time until an accepted result. Compare a latency-sensitive task separately from work that can wait. MustHave has not reproduced Fireworks’ speed or savings claims.

The operational controls may matter as much as throughput for a shared team account. Budget ownership and usable audit exports should be part of the evaluation alongside model routing. Follow our AI news coverage for related infrastructure releases.