DeepSeek V4.1 Flash

A frontier-class generalist at flash-tier prices: a million tokens of context, strong agentic coding, and output fast enough to sit in an interactive loop.

DeepSeek V4.1 Flash is the efficiency cut of the V4 family: the same million-token context as its larger siblings, with aggressive KV-cache compression and speculative decoding doing the work of keeping it fast and cheap. We serve it with TP4 and DSpark speculative decoding.

It benchmarks like a frontier model on knowledge and reasoning, and holds the top of its class on agentic coding — which makes it the default choice when you want one model to cover chat, code, and long-document work before specialising it on your own traffic.

deepseek/deepseek-v4.1-flash

What’s it good for

Agentic coding

Strong on DeepSWE v1.1 and Terminal-Bench 2.1. A capable base for coding agents before post-training sharpens it on your repository.

Long-context work

A million tokens of context at a price that lets you actually use them — whole codebases, document sets, or long agent histories in one window.

General chat and reasoning

Frontier-class MMLU and GPQA scores. A sensible default when a workload has not yet earned a specialist.

Benchmarks

Figures from the DeepSeek-V4.1-Flash model card, October 2026.

Terminal-Bench 2.190.6pass@1

GPQA Diamond90.9pass@1

DeepSWE v1.174.2resolved

MMLU-Pro74.15-shot, base model

Pricing & Specs

Latency is a choice, not a property of the model — pick the service tier your workload needs.

Input / 1M: $0.30
Cached input / 1M: $0.006
Output / 1M: $1.20
Context: 1M

Weights: Open — deepseek-ai/DeepSeek-V4.1-Flash
Released: September 2026
Quantization on Gradiated: FP4
Max output: 131K tokens
Streaming: Yes
Tool calls: Parallel tool calls, reasoning effort

Getting Started

Point your existing client at our base URL and name the model. Your traffic starts accruing as traces — the raw material post-training is built from.

from openai import OpenAI
import os

client = OpenAI(
    base_url="https://api.gradiated.com",
    api_key=os.environ["GRADIATED_API_KEY"],
)

response = client.chat.completions.create(
    model="deepseek/deepseek-v4.1-flash",
    messages=[{"role": "user", "content": "Summarize this document."}],
)

Read the quickstart

Hardware

What DeepSeek V4.1 Flash runs on — our serving fleet first, then what the open weights support anywhere else.

NVIDIA B200 / B300: What Gradiated serves on — four per replica, TP4 at FP4
NVIDIA H100 / H200: Self-host across a multi-GPU node, FP8
Inference engines: vLLM and SGLang