Fireworks 公佈 Ultra-fast Agent 推理預覽與 Sovereign 部署方案

Fireworks announced a slate of enterprise AI products on November 3, 2026, at its first annual Forge conference, led by an Ultra-fast inference preview that it says reached up to 1,500 tokens per second on DeepSeek V4.1 Flash, alongside tools for model routing, evaluation, specialized serving and a private preview of Fireworks Sovereign for customer-owned infrastructure.

Matching Inference Speed and Cost to the Workload

Ultra-fast is in public preview on NVIDIA Blackwell GPUs and is designed for multi-step agents, where each model call adds to the wait for a result. Fireworks reported that completion time for its 30-turn agent benchmark fell from one minute to eight seconds with Ultra-fast. The capability is aimed at time-sensitive work, with cheaper options for jobs that can wait.

For delay-tolerant work, Background Mode runs eligible tasks, such as evaluation-data generation and overnight research, at half price. Multi-day Prompt Caching targets the cost of repeated context. Higher serverless rate limits and Elastic Burst are generally available for busier workloads; Elastic Burst can exceed a customer’s normal usage ceiling when spare capacity exists, without code or pricing changes.

A September 14, 2026 Fireworks benchmark of DeepSeek V4.1 Flash shows why repeated context matters. In its DeepSWE run, 99.6% of input tokens were cache hits, and cache-hit charges represented 60% of the bill. That breakdown makes repeated context a material cost target for long-running agents, rather than treating generated output as the only source of spending.

The same cost theme extends to model specialization. Ember-1, introduced on September 23, 2026, as a serverless Research Preview, used about 40% fewer tokens than Kimi K3 at comparable quality in Fireworks’ evaluations. It is available on serverless and dedicated deployments.

Nexus Desktop and Subscription Pricing

Fireworks Nexus routes across open and closed models inside existing coding and coworking tools, using personalized model selection and cache awareness. It also centralizes spending visibility and controls for organizations. Nexus Desktop launched alongside the existing Nexus CLI, with Nexus running in the Codex app, Claude Desktop and OpenCode.

New Nexus Pro plans start at $8 per month, while Teams plans start at $10 per seat per month. Enterprise plans add single sign-on, governance, shared budgets and access controls.

The Nexus Desktop documentation lists a universal macOS build for Apple silicon and Intel Macs, with Linux and Windows support coming soon. The CLI runs on macOS and Linux, with limited Windows support through Git Bash. Desktop connections include ChatGPT, Claude Code, Cursor IDE, VS Code and OpenCode, while Claude Desktop support is described as limited.

Evaluation and Specialized Model Serving

No Eval, now in private preview, uses traces from production systems and signals from products to build evaluations for specific tasks. The evaluations are intended to expose model weaknesses and show whether later training resolves them, addressing work that can otherwise require weeks of data labeling and grading.

Specialized Model Inference entered public preview for serving LoRA adapters on Fireworks’ shared serverless pool, initially with GLM 5.3 Flash. Customers pay per token at the base model’s serverless price without creating a deployment. Loading an adapter, keeping it loaded and idle time carry no serving charge, while training is billed separately.

For GLM 5.3 Flash, the documented rates are $0.15 per million input tokens, $0.03 per million cached input tokens and $0.50 per million output tokens.

The preview supports a maximum LoRA rank of 64 and up to three loaded adapters per account for the base model. Fireworks documents loading at under five minutes, a first request that can take up to one minute and subsequent requests at approximately base-model latency, with no latency guarantees during the preview.

Adapters run on GPUs shared with other customers’ adapters. Dedicated deployments are the recommended option for workloads needing consistent latency, isolated GPUs, a specific region or higher-precision serving; serverless is aimed at experimentation and spiky traffic.

Enterprise Controls and Sovereign Deployment

Model governance, spending controls and exportable audit logs are now generally available. Centralized account management, regional data controls and customer-managed encryption remain in preview.

Fireworks Sovereign, in private preview, runs the company’s inference engine in a customer’s cloud, neocloud or data center. Inference and training stay inside the customer’s environment and under its security and change-control policies. Fireworks reports managing more than 25,000 GPUs in customer environments.

The operating model divides responsibility between the two parties. Fireworks qualifies the cluster, deploys models and handles upgrades, monitoring, performance tuning and incident response, while the customer retains ownership of the architecture and policies. Customers manage ingress, load balancing, DNS, TLS certificates and approved outbound routes. Fireworks’ control plane uses a dedicated auditable credential, with outbound traffic limited to model-weight and image pulls plus operational telemetry.

Fireworks says Sovereign delivers 20% to 50% more throughput on identical hardware and about 2.3 times the tokens per hardware dollar. Optional hybrid overflow moves eligible workloads to Fireworks Cloud in customer-approved regions when a cluster saturates, then returns them to customer GPUs as demand falls, using the same API.

The Forge lineup makes model choice, inference economics and infrastructure control part of the same enterprise decision. The central qualification is that each option carries different operating conditions: serverless adapters avoid dedicated capacity yet run on shared GPUs without preview latency guarantees, while Sovereign combines customer-owned infrastructure with a serving stack operated by Fireworks.