Fireworks uses its first Forge conference to launch 1,500 token-per-second inference and run its engine on customers' own GPUs

The headline is Ultra-fast, now in public preview on NVIDIA Blackwell GPUs. Fireworks says it reaches up to 1,500 tokens per second, as demonstrated on DeepSeek V4.1 Flash, and that in a benchmarked 30-turn agent task completion time dropped from 1 minute to 8 seconds[1]. That matters for agents because every model call in a chain adds to the wait.

For work that can wait, Background Mode runs eligible jobs, such as eval data generation or overnight research, at half price. Multi-day prompt caching cuts the cost of agents that resume after a pause and would otherwise reprocess the same context, and higher serverless rate limits plus Elastic Burst, both generally available, let teams run more agents at once[1].

Fireworks also pointed to its own Ember-1 model, which it says used about 40% fewer tokens than Kimi K3 at comparable quality in its evaluations[1].

Nexus, which mixes open and closed models with per-person model selection and central spend tracking, now has a desktop app alongside the Nexus CLI. It runs in the Codex app, Claude Desktop and OpenCode. Pro plans start at $8 a month and Teams plans at $10 per seat a month, with enterprise plans adding SSO, shared budgets and access controls[1].

No Eval, in private preview, turns production traces and product signals into task-specific evaluations, aiming to remove weeks of labeling and grading before a team can tell whether training helped. Specialized Model Inference, in public preview, serves LoRA adapters serverlessly with per-token pricing and no charge while idle, starting with GLM 5.3 Flash. Fireworks is also offering free training on Muse Glimmer 30B through November 30[1].

For enterprises, model governance, spending controls and exportable audit logs are now generally available, with regional data controls and customer-managed encryption in preview. The bigger move is Fireworks Sovereign, a private preview that runs Fireworks' inference engine on a customer's own cloud, neocloud or data center, inside their security perimeter. Fireworks says it already manages over 25,000 GPUs in customer environments[1][3].

Separately, Fireworks said it will expand capacity in Europe with NVIDIA Vera Rubin and Grace Blackwell systems, with deployments starting in 2027[2].

Open-weight models have closed much of the quality gap, so the fight is shifting to speed, cost and control. Fireworks is betting that enterprises will pay for fast agents, cheap background work and the option to keep everything on hardware they govern, rather than depend on a frontier lab's pricing and model retirements. All of the speed and efficiency numbers here are Fireworks' own, so independent benchmarks will be worth watching as the previews open up.