How I keep agent loops fast without waiting on slow models

Agentic and interactive coding apps lose their rhythm when model generation is the slowest step in the loop. Vercel AI Gateway now serves a DeepSeek Flash Ultrafast endpoint through Fireworks that generates upwards of 1,000 tokens per second with a 1M-token context window.

What it is

A public preview model on AI Gateway, served through Fireworks' accelerated serverless endpoint, with no waitroom or session time limit. You call it by setting deepseek/deepseek-v4.1-flash-1017-ultrafast as the model name.

Applied case

A developer building an interactive coding agent needs the app to generate output and move to the next step quickly. They install the latest Vercel CLI, run setup, and select the model in Claude Code, Codex, or Cursor. With a 1M-token context window, the agent can work over large codebases and long inputs in a single pass, and usage and cost are tracked through the Gateway, which charges no platform fee on inference, including BYOK requests.

• Call the Ultrafast Flash endpoint by model name for agent loops needing upwards of 1,000 tokens per second on large

When to use it

Use when your agent needs fast token generation, a huge context window, and unified gateway tracking. Skip when you need a model outside the DeepSeek Flash family or the public preview status is a blocker for production guarantees.

Read more

Vercel changelog: DeepSeek Flash Ultrafast

AI Gateway coding agents guide