Fireworks AI, the company that hosts open-weight models for developers, unveiled a batch of products at its first Forge conference, led by a serving tier it says reaches 1,500 tokens per second. The company described the launches on its own blog. Its pitch is that businesses should own AI tuned to their work instead of renting it from a few closed labs.
A token is a small chunk of text, usually a short word or a piece of one, that a model reads and writes. Tokens per second is how fast the AI types its answer. At 1,500, a long reply lands almost at once. Fireworks says it measured that peak on DeepSeek V4.1 Flash, running on NVIDIA Blackwell chips. The tier, called Ultra-fast, is in public preview.
Speed matters most for agents, the AI programs that make many model calls to finish one job, because every wait stacks on the last. In a 30-turn agent task the company benchmarked, Fireworks says completion dropped from one minute to eight seconds. That test is Fireworks’ own, and the post cites no outside verification.
The other half of the launch is about paying less for work that can wait. Background Mode runs eligible jobs at half price, such as generating test data or running research overnight. Prompt caching that lasts several days cuts the cost of resending the same context, and higher rate limits let more agents run at once. Fireworks also introduced Ember-1, its own model, which it says used about 40 percent fewer tokens than Kimi K3 at comparable quality in its evaluations.
Nexus tackles a different problem: staff who want more model choice without leaving the tools they already use. It plugs open and closed models into OpenCode, Claude Desktop, and the Codex app, next to existing subscriptions, and picks a model for each task. Finance teams get a dashboard of spend by model. Fireworks lists Pro plans from $8 a month and Teams plans from $10 per seat per month, with SSO, shared budgets, and access controls on the enterprise tier.
For companies that want to train their own models, Fireworks put two pieces in preview. No Eval turns production traces into custom tests, which the company says normally take weeks of labeling and grading to write. Specialized Model Inference serves a trained LoRA adapter (a small add-on that tunes a base model) with per-token pricing and no charge while idle, beginning with the GLM 5.3 Flash model. Fireworks is also offering free training runs on its Muse Glimmer 30B model until November 30.
The enterprise block is the most telling. Rules for who can use which model, budget limits, and audit logs that can be exported are now generally available. Fireworks Sovereign, in private preview, installs the company’s serving engine in a customer’s own cloud or data center. Fireworks says the technology can more than double the output a GPU dollar buys, and that it already runs more than 25,000 GPUs inside customer sites. For regulated buyers, the draw is independence: if a vendor raises prices or retires a model, their roadmap does not change.
Taken together, the launches try to make open models feel like a complete product with speed, cost controls, and compliance included, which is the ground closed labs have long held. Until Fireworks publishes a rate for Ultra-fast, the eight-second result is a speed figure without a price, and the price decides whether an agent workflow can afford it.
Reported by the Fireworks AI blog, from its Forge conference announcements (no publication date given in the source).
