Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Coding tasks

DeepSeek V4.1 Flash, ultrafast on Vercel

Written November 8, 2026

Vercel's AI Gateway serves DeepSeek V4.1 Flash 1017 Ultrafast in public preview with a 1M context. The vendor speed claim, and the check.

What it is

DeepSeek V4.1 Flash 1017 Ultrafast is now available on Vercel's AI Gateway through Fireworks, in public preview with no waitroom or session time limit. The model ID is deepseek/deepseek-v4.1-flash-1017-ultrafast. It serves the October 17 update of DeepSeek V4.1 Flash through Fireworks' accelerated serverless endpoint. That is the whole change: this is not new base-model weights. The October 17 checkpoint is the same model, served faster. It has a 1M-token context window for large codebases and long inputs, and Vercel says it is suited to interactive coding, reasoning, and agentic workflows where applications need to generate output and move to the next step quickly.

To use it in Claude Code, Codex, Cursor, and more, install the latest Vercel CLI and run npx vercel ai-gateway setup, then select the model in the agent. AI Gateway provides one API for calling models, tracking usage and cost, and configuring retries, failover, and performance tuning. It reflects provider pricing with no markup and charges no platform fee on inference, including on Bring Your Own Key requests.

The speed claim, and the check

Vercel's changelog says the model can generate upwards of 1,000 tokens per second. That is a vendor figure, from the Fireworks endpoint, not a measurement anyone independent has published. Vercel's own model page lists the same model at 724 tokens per second, with a 0.7-second latency figure. The two numbers come from different measurements, and neither is your workload. Treat the 1,000 figure as the vendor's ceiling, and the 724 figure as the Gateway's own listing, and measure the one that matters: tokens per second on your prompts, in your harness.

The price is on the same model page: $2.40 per million input tokens and $9.60 per million output tokens, through Fireworks. For comparison, the Fast variant of the same checkpoint on the same page costs $0.60 per million input and $2.40 per million output. Ultrafast costs four times as much per token as Fast. Whether it is worth that depends on what the extra speed buys on your task, and the only way to know is to run the same task on both and compare tokens per second, cost, and quality.

What it is for

Ultrafast fits the same shape as any fast, cheap model: interactive coding, where waiting on output breaks flow, and agentic workflows, where each step waits on the last model's output. The 1M-token context window matters here too. A coding agent working across a large codebase can hold more of it in context, and a long agentic run can keep more history before it needs to compact. The model also supports text and image input, configurable reasoning, tool calling, structured outputs, and automatic prefix caching, per the Gateway model page.

Weaknesses and limits

The check to run

Take one real interactive coding task, the kind you run every day, and run it through the Gateway on Ultrafast and on the Fast variant, with the same prompt. Record tokens per second, total cost, and whether the output was good enough to use without rework. If Ultrafast finishes meaningfully faster at a cost you accept, switch the interactive path to it. If the quality is the same and the speed gain is small, the Fast variant is the better default. Do not switch on the vendor's number alone.

Limits

Related playbooks

Sources

Back to: Coding playbooks hub · Also: Foundry Local runs Qwen3.8-27B on your GPU · Model prices: what the major models cost · Model specs: context, tools, and cutoffs in one table