DeepSeek V4.1 Flash 1017 Ultrafast brings 1,000+ tokens/sec to Vercel AI Gateway

DeepSeek's ultrafast Flash model arrives on Vercel AI Gateway

Vercel has made DeepSeek V4.1 Flash 1017 Ultrafast available through AI Gateway, its unified API for language models. According to the Vercel blog, the model can generate more than 1,000 tokens per second, with Fireworks supplying the accelerated serverless endpoint behind it.

A speed-first variant of DeepSeek's Flash model

The Ultrafast endpoint serves the October 17 update of DeepSeek V4.1 Flash. Vercel describes it as a fit for interactive coding, reasoning and agentic workflows — situations where an application needs to finish generating output and move on to its next step rather than wait on a slow stream.

The model also carries a 1M-token context window, which Vercel frames as suited to working with large codebases and long inputs. It is currently in public preview, and per the announcement there is no waitroom and no session time limit, so developers can experiment without queuing or being cut off mid-run.

How to switch to it

On AI Gateway the model is addressed as deepseek/deepseek-v4.1-flash-1017-ultrafast. To use it inside coding agents such as Claude Code, Codex or Cursor, Vercel says to install the latest Vercel CLI and run:

npx vercel ai-gateway setup

After setup, the model can be selected directly in the agent. Vercel also offers it in the model playground for anyone who wants to gauge the speed before wiring it into a project.

What the gateway layer adds

AI Gateway is the connective tissue here. Vercel pitches it as a single API for calling models, tracking usage and cost, and configuring retries, failover and performance optimizations that the company says deliver higher uptime than calling a provider directly. The service also includes built-in custom reporting, Zero Data Retention support, per-API-key budgets and routing rules.

On pricing, Vercel says the gateway passes provider pricing through without markup and takes no platform fee on inference — including on Bring Your Own Key requests, where developers route their own provider credentials through the platform.

Why it matters

Throughput has quietly become a first-order constraint in AI application design. A single slow completion is tolerable; an agent making dozens of sequential model calls is not, because latency compounds across every step. A model generating at over 1,000 tokens per second changes what feels practical for interactive coding assistants and multi-step agents, where a long answer can otherwise take tens of seconds to finish streaming.

The packaging matters as much as the raw speed. By serving the endpoint through its gateway with retries, failover and usage tracking, and by charging no inference markup or platform fee, Vercel is making a low-friction play for developers building latency-sensitive applications: point an existing agent at the model string and keep observability in one place.

Caveats remain. The offering is a public preview, real-world token rates can vary with workload, and Vercel's figure describes the endpoint's capability rather than a guaranteed floor. Even so, for teams whose agents spend most of their wall-clock time waiting on generation, this kind of endpoint is a straightforward way to shrink that wait.