Bay Area AI company Fireworks unveils faster, cheaper ways for businesses to run open AI models
Fireworks, founded by former Meta PyTorch engineers, announced the products at its first Forge conference.
Fireworks AI, a Bay Area company that runs AI models for other businesses, announced a set of new products Tuesday at Forge, its first yearly conference.
Fireworks does what the industry calls inference. That is the work of running an already-trained AI model to answer questions or carry out tasks. Instead of paying OpenAI or Anthropic for their closed models, companies can use Fireworks to run open models, ones whose code is freely published, such as those from the Chinese lab DeepSeek. According to its website, Fireworks was co-founded by CEO Lin Qiao and a team with experience from PyTorch, the AI software toolkit created at Meta, and from Google.
The biggest launch is about speed. AI agents, programs that work through a task step by step, may call a model dozens of times, and each call makes the user wait. Fireworks said its new Ultra-fast option, running on Nvidia's newest Blackwell chips, can produce up to 1,500 tokens, or chunks of text, per second. It said a 30-step agent task that took one minute finished in eight seconds.
For work that can wait, such as overnight research, Fireworks introduced Background Mode at half the normal price.
It also launched Nexus Desktop, which lets workers switch among AI models inside tools they already use, such as Anthropic's Claude Desktop and OpenAI's Codex app. Plans start at $8 a month for individuals and $10 per person per month for teams.
For banks and other companies that cannot send data outside their walls, Fireworks Sovereign, in early testing, runs Fireworks' software on a customer's own computers. The company said it already manages more than 25,000 GPUs, the chips that run AI, inside customers' facilities. The speed and cost figures are Fireworks' own.
Source: Fireworks AI announcement, Nov. 3, 2026; Fireworks company website