Fireworks strengthens AI inference and deployment control

Fireworks introduced a series of enterprise AI tools around inference, routing, and cost control. The vendor also offers a Sovereign solution for clients who want to keep their data and models on their own infrastructure.

Fireworks unveiled on November 3, 2026, during its first Forge conference, a new wave of AI products for businesses. The highlight is Ultra-fast, a public preview inference mode that, according to the company, reaches up to 1,500 tokens per second on DeepSeek V4.1 Flash. Behind the speed demonstration, the challenge is broader: reducing wait times, lowering costs, and maintaining control over infrastructure—three critical issues in AI production deployments.

When inference becomes a budget issue

Ultra-fast targets multi-step agents, systems that chain several model calls before delivering a response. Fireworks claims to have reduced the completion time of its agent benchmark over 30 rounds from one minute to eight seconds. The service is in public preview on NVIDIA Blackwell GPUs. Meanwhile, the company reserves other modes for less urgent tasks: Background Mode, intended for deferrable processes like evaluation data generation or overnight search, is billed at half price for eligible uses.

Pricing logic also involves caching. Multi-day Prompt Caching aims to limit the cost of repeated contexts, a factor that quickly becomes significant in long-running agents. Fireworks cites a benchmark from September 14, 2026, on DeepSeek V4.1 Flash: in its DeepSWE run, 99.6% of input tokens were cache hits, and these loads accounted for 60% of the bill. In other words, previously seen text sometimes weighs more than generated text. It’s less glamorous than launching a model but much more useful for the finance department.

Routing, evaluation, and specialized models: the toolbox expands

The conference also introduced Nexus, a routing layer between open and closed models integrated into coding and coworking tools already used by teams. Fireworks adds personalized model selection, cache handling, and a centralized spending view. Nexus Desktop comes with the existing CLI, and the service integrates with Codex, Claude Desktop, and OpenCode. Nexus Pro plans start at 8 dollars per month, Teams offers at 10 dollars per seat per month, while Enterprise plans add single sign-on, governance, shared budgets, and access controls.

Another component, No Eval, in private preview, relies on production traces and product signals to build targeted evaluations. The idea is to measure concrete weaknesses without weeks of manual labeling. Fireworks also launched Specialized Model Inference in public preview to serve LoRA adapters on its shared fleet, starting with GLM 5.3 Flash. Documented pricing for this model is 0.15 dollars per million input tokens, 0.03 dollars per million cached input tokens, and 0.50 dollars per million output tokens.

Businesses want control, not just promises

In the enterprise arena, Fireworks emphasizes governance. Model control, budgets, and exportable audit logs are now generally available. Other features remain in preview, such as centralized account management, regional data controls, and client-managed encryption. This positioning addresses a simple need: to integrate generative AI into existing operational rules without turning each project into a parallel undertaking.

The heaviest piece of the lot is called Fireworks Sovereign. This private offering runs the vendor’s inference engine in the cloud, a neocloud, or a data center owned by the client. Inference and training then remain within the client’s environment, under their security and change policies. Fireworks says it manages more than 25,000 GPUs in client environments. The company handles cluster qualification, model deployment, updates, monitoring, performance optimization, and incident response, while the client retains control over architecture, policies, ingress, load balancing, DNS, TLS certificates, and authorized outbound routes.

Fireworks claims Sovereign delivers 20% to 50% more throughput on identical hardware and about 2.3 times more tokens per hardware dollar. A hybrid overflow option can also redirect eligible loads to Fireworks Cloud, in client-approved regions, when the cluster saturates, then revert to client GPUs when demand drops. Fundamentally, the dividing line is clear: on one side, server options without dedicated capacity but no latency guarantees during preview; on the other, a stack operated by Fireworks on infrastructure owned by the client. CIOs rarely have a spontaneous taste for binary choices, but they love having the trade-offs explained.

Key points

  • Ultra-fast reaches up to 1,500 tokens per second.
  • Agent benchmark: 30 rounds reduced from one minute to eight seconds.
  • DeepSWE: 99.6% cache hits, according to Fireworks.
  • Nexus Pro starts at 8 dollars per month.
  • Fireworks manages more than 25,000 client GPUs.
  • Sovereign claims up to 2.3 times more tokens per dollar.

By the numbers

  • 1,500 tokens per second — peak announced by Fireworks on DeepSeek V4.1 Flash, November 3, 2026.
  • 99.6% — share of input tokens cached in DeepSWE, benchmark of September 14, 2026.
  • 60% — share of the bill represented by cache-hit loads on the same run.
  • 25,000 GPUs — fleet managed by Fireworks in client environments.
  • 8 dollars per month — entry price for the Nexus Pro plan.

To read