Vercel Makes DeepSeek V4.1 Flash Ultrafast Available with 1,000 Tokens per Second Generation
Chinese model arrives at Vercel's AI Gateway via Fireworks with a 1 million token context window and speed geared toward interactive coding workflows and autonomous agents
Vercel now offers DeepSeek V4.1 Flash 1017 Ultrafast in its AI Gateway, a unified layer for accessing language models. The model is served through Fireworks, which operates the accelerated serverless endpoint responsible for generation speed.
According to the company, the Ultrafast version delivers over 1,000 tokens per second. The performance is aimed at interactive coding, reasoning, and agentic workflows—scenarios where the system needs to produce output and quickly move to the next step.
What speed changes in practice
The throughput gain alters the experience in assisted development tools. Instead of waiting for long blocks of code to complete, the developer receives the response in a continuous stream and can iterate on it almost in real time.
The model corresponds to the October 17 update of DeepSeek V4.1 Flash. Distribution via Fireworks uses the partner's accelerated inference infrastructure, which Vercel integrates into its gateway without passing on additional platform costs.
1 million token context
The 1 million token context window allows processing extensive codebases and long inputs in a single call. In practice, this reduces the need to fragment files or split reasoning into multiple manual steps.
The feature is particularly relevant for agents that need to maintain a history of decisions and code snippets across prolonged tasks without losing coherence between interactions.
How to access
Availability is in public preview, with no waitlist or session time limit. Access is granted by setting the identifier deepseek/deepseek-v4.1-flash-1017-ultrafast in API calls.
Vercel has also integrated the model into coding tools such as Claude Code, Codex, and Cursor through the Vercel CLI configuration command. The AI Gateway works as a unified API for calls, usage and cost tracking, as well as configuring retries, failover, and performance optimizations.
The platform states that it reflects provider prices without additional margin and does not charge a fee on inference, including requests with your own key (BYOK). The announcement reinforces the race for inference speed as a competitive differentiator in the language model market.