Fireworks Forge adds Nexus Desktop and faster inference, with tradeoffs

Fireworks used its November 3 Forge conference to expand model routing, faster inference and customized model serving. The practical choice is which part of the stack to adopt. Nexus Desktop connects existing AI tools to model selection, Ultra-fast trades a higher token bill for speed, and Specialized Model Inference removes the need to reserve a deployment for a supported adapter.

The Forge announcement mixes available products with public and private previews. Prices and documentation below were checked on November 5, 2026.

Nexus Desktop adds a simpler setup, with platform limits

Nexus Desktop currently runs on macOS, with a universal build for Apple silicon and Intel Macs. Windows and Linux desktop versions are still planned. The CLI runs on macOS and Linux. Windows support is limited to Git Bash.

The desktop app provides switches for connecting tools. After enabling a tool and restarting it, Nexus Auto appears in its model picker. Disabling the switch and restarting restores the original setup. Desktop and CLI share their settings, so switching between them does not create a separate configuration.

Fireworks lists Claude Code, Codex, ChatGPT, Cursor, VS Code and OpenCode among supported tools. Its harness documentation explicitly labels Claude Desktop support as limited on macOS. A supported app name therefore should not be read as a promise that every workflow behaves identically.

What the $8 Nexus plan actually includes

The Nexus plan guide lists Pro at $8 per month and Team at $10 per seat per month. Team supports up to 200 seats. Enterprise uses usage-based pricing and a sales process rather than a published flat subscription.

Pro and Team include a Fireworks usage allowance, with rolling and monthly resets. Anthropic and OpenAI turns use the customer’s own provider account. Those turns do not become free because Nexus routed them. Fireworks turns pause when the included allowance is exhausted, until the relevant window resets.

There is also a model-selection distinction. Pro and Team can route to open models through Nexus Auto, but do not include general direct access to the open-model catalog. Enterprise adds that access. Nexus accounts, billing and keys are separate from the Fireworks platform used for inference and training.

The Auto documentation describes choosing an open, Anthropic or OpenAI model for each turn. It exposes the serving model in the response, giving developers a way to inspect what actually handled a request. A useful pilot should compare the combined provider bill and accepted output, rather than treating the Nexus subscription price as the entire cost.

Ultra-fast is a speed purchase

Ultra-fast is a public preview. Fireworks reports up to 1,500 tokens per second and an eight-second completion time for its demonstrated 30-turn agent task, versus one minute in the comparison. These are vendor results, not ByteForward measurements.

The mode guide currently identifies DeepSeek V4.1 Flash 1017 as the supported Ultra-fast model. Switching uses its dedicated router identifier. Fireworks describes this as a serving mode, rather than a different underlying model, and warns that preview behavior can change quickly.

The published rates make the tradeoff concrete. Per million tokens, standard Flash 1017 costs $0.30 for input, $0.006 for cached input and $1.20 for output. Ultra-fast costs $2.40, $0.048 and $9.60 respectively. Each listed rate is eight times the standard rate.

For a hypothetical workload totaling one million uncached input tokens and 100,000 output tokens, those rates produce $0.42 in standard token charges or $3.36 in Ultra-fast token charges. This calculation holds token use constant and excludes other costs. Faster completion can still be worthwhile when someone is waiting, but the relevant measure is cost per usable result.

That is the same practical issue explored in ByteForward’s Flash 1017 demonstrations and the cost of waiting. Tool execution, retries and review time can dominate a complete job even when model generation gets faster.

Serverless adapters avoid idle deployment charges

Specialized Model Inference, also in public preview, serves LoRA adapters on shared GPUs. Initial support is limited to GLM 5.3 Flash, a maximum LoRA rank of 64 and three loaded adapters per account for that base model.

Serving is charged at the base model’s token rates, with no fee for loading an adapter, keeping it loaded or leaving it idle. Training is billed separately. This makes a small experiment easier to budget than a continuously reserved deployment.

There are operational limits. The first request after loading can take up to a minute, and Fireworks provides no preview latency guarantee. Training also uses higher numeric precision than serving, so teams should compare task-level results on held-out examples. The documentation points customers needing consistent latency, isolated GPUs or higher-precision serving toward dedicated deployments.

Available capacity is not reserved capacity

Higher serverless limits and Elastic Burst are generally available. The rate-limit documentation says Burst can temporarily raise eligible accounts above their normal ceiling when spare capacity exists. It applies automatically without a separate premium, but the extra capacity is not guaranteed. Applications still need to handle rate limits and overloaded responses.

The remaining Forge announcements need different expectations. No Eval, which turns production signals into task-specific evaluations, and Fireworks Sovereign are private previews. Background Mode was announced at half price for eligible work that can wait. Teams should verify eligibility and execution details before building a budget around it.

Sovereign targets inference managed by Fireworks on infrastructure the customer controls. That is a different deployment decision from adding Nexus to an employee’s desktop or serving an adapter on shared infrastructure.

Start with the bottleneck you can measure

A sensible evaluation starts with one workload. For an interactive agent, compare elapsed time and the full bill across standard and Ultra-fast. For a customized model, check quality and cold-start behavior before committing to serverless adapters. For employee tooling, test Nexus against the subscriptions and app features people already use.

Forge broadens the choices, but those choices solve different problems. The strongest adoption case is a measured improvement in the work a team needs to finish, with the launch stage, limits and total cost kept visible.

Featured illustration by ByteForward using the official Fireworks logo for independent editorial identification. No affiliation or endorsement is implied.