Vercel Is First to Plug In DeepSeek at 1000 Tokens/Sec

Vercel's AI Gateway has launched DeepSeek V4.1 Flash 1017 Ultrafast. The official core numbers are straightforward: generation speed exceeds 1,000 tokens/s. This version corresponds to the October 17 update of DeepSeek V4.1 Flash and is provided through Fireworks' accelerated serverless endpoint.

Both Vercel Developers' launch post and the changelog mention that it targets interactive coding, reasoning, and agent workflows. In everyday terms: when you are writing code and waiting for completions, asking the model to explain a piece of logic continuously, or having an agent call tools step by step, output speed directly affects the experience. The 1M-token context window is another selling point, suitable for large codebases and long inputs. During the public beta, there is no waiting room and no session time limit.

The model name is deepseek/deepseek-v4.1-flash-1017-ultrafast. In the AI SDK, you can call it like this:

import { streamText } from 'ai';

const result = streamText({
  model: 'deepseek/deepseek-v4.1-flash-1017-ultrafast',
  prompt: 'Write a TypeScript helper that retries a failed request with backoff.',
});

If you use tools such as Claude Code, Codex, or Cursor, you can install the latest Vercel CLI and run npx vercel ai-gateway setup, then select this model in your agent. Model details and the playground are on Vercel's AI Gateway page.

AI Gateway itself is Vercel's unified model invocation layer, providing a unified API, tracking usage and costs, and configuring retries, failover, and performance optimization. The official statement says it follows provider pricing, adds no markup, and charges no inference platform fee. The most common question in the comments is about pricing. One commenter did the math using $2.4 for input and $9.6 for output, saying it is 16 times more expensive than DeepSeek's official service. This claim points to the price difference between the Fireworks accelerated endpoint and the official API; actual billing is subject to the model page.

The speed numbers are impressive, but real experience still depends on load, network, and task type. Under the announcement, a commenter joked that Ultrafast had already finished while another model had not even reached the caching part. For people building agents and interactive coding, this speed is worth trying. Whether the price is worth it depends on your call volume and latency requirements.