Updated November 18, 2026

Uncensored DeepSeek V4.1 API.

Every uncensored and abliterated DeepSeek V4.1 Flash you can call by API, with hosts, prices and our measurements, and how each compares with the original model.

Short answer

You can run one on OpenUncensored, which serves 8 uncensored DeepSeek V4.1 Flash builds through one OpenAI and Anthropic compatible API (https://api.openuncensored.com/v1), from Venice, heabsy, NanoGPT, Lebrel, Refuseless and UncensuredAI, with automatic fallback to the next host when one fails and no account to create. The best one on our tests is DeepSeek V4.1 Flash 1017 Uncensored (deepseek-v4.1-flash-1017-uncensored): ranked #1 of 34 on the OpenUncensored Score (69.8), answers 100% of 36 requests censored models refuse and runs at about 239 tok/s, from $0.33 per million input and $1.32 per million output tokens.

On the API

Every uncensored DeepSeek V4.1 Flash you can call.

ModelModel idTypeRank (score)AnswersAgent tasksSpeed$ / 1M in · outContext
DeepSeek V4.1 Flash 1017 Uncensoreddeepseek-v4.1-flash-1017-uncensoredUncensored fine-tune#1 (69.8)100%2/5239 tok/s$0.33 · $1.321M
DeepSeek V4.1 Flash 0910 Uncensoreddeepseek-v4.1-flash-0910-uncensoredUncensored fine-tune#4 (66.1)97%2/5209 tok/s$0.33 · $1.32512K
DeepSeek V4.1 Flash Uncensoreddeepseek-v4.1-flash-uncensoredUncensored fine-tune#5 (64.3)100%1/576 tok/s$0.12 · $0.451M
DeepSeek V4.1 Flash Hereticdeepseek-v4.1-flash-1017-hereticHeretic (automated abliteration)not ranked (72.7)94%–53 tok/s$0.20 · $0.79500K
DeepSeek V4.1 Flash 0910 Uncensored Thinkingdeepseek-v4.1-flash-0910-uncensored-thinkingUncensored fine-tune, reasoningnot scored yet––199 tok/s$0.33 · $1.32512K
DeepSeek V4.1 Flash 1017 Uncensored Thinkingdeepseek-v4.1-flash-1017-uncensored-thinkingUncensored fine-tune, reasoningnot scored yet––260 tok/s$0.33 · $1.321M
DeepSeek V4.1 Flash Heretic Thinkingdeepseek-v4.1-flash-1017-heretic-thinkingHeretic (automated abliteration), reasoningnot scored yet––106 tok/s$0.20 · $0.79500K
DeepSeek V4.1 Flash Uncensored Thinkingdeepseek-v4.1-flash-uncensored-thinkingUncensored fine-tune, reasoningnot scored yet––228 tok/s$0.33 · $1.321M
Answers: share of 36 requests that censored models refuse. Agent tasks: multi-step tool tasks finished out of 5. Prices are the cheapest host's USD per million tokens. Scores measured November 8, 2026; reasoning variants are not scored separately.

1017 and 0910 are later V4.1 Flash checkpoints, named by date; the id without a date is built on the base V4.1 Flash release. Thinking variants run the same weights with reasoning switched on.

Which one

Which build to use.

Hosts

Where DeepSeek V4.1 Flash 1017 Uncensored runs.

HostInputCached inputOutputNative tool calls
Venice$0.33$0.008$1.32yes
heabsy$0.40$0.04$1.40yes
NanoGPT$0.33–$1.32no
Lebrel$0.33$0.01$1.32yes
USD per million tokens, live from the router. One key reaches every host; a failed request moves to the next one, and requests with tools go to hosts with native tool calls first. Send "routing": {"strategy": "cheapest"} to try the cheapest host first.

Uncensored vs original

Does uncensoring hurt it?

We ran an uncensored DeepSeek V4.1 Flash build, dealignai/DeepSeek-V4.1-Flash-1017-UNCENSORED-FP8, and the original, deepseek-ai/DeepSeek-V4.1-Flash-1017, through our open Uncensored Agent Eval: 110 public tasks on Hugging Face, the same harness and the same grader for both.

TaskOriginalUncensoredChange (points)
Answers legitimate requests97%100%+3
Picks the right next agent step80%73%-7
Breaks out of a failing loop85%75%-10
Share of items graded correct. Gaps under about 15 points are within noise at this size. The builds hosts serve under the ids above may differ from the one measured here.

Quickstart

Call it in a minute.

Buy a prepaid key on the credits page with no account (the key is the account), or point your agent at agents.md and let it buy one.

from openai import OpenAI

client = OpenAI(base_url="https://api.openuncensored.com/v1", api_key="ou-...")

r = client.chat.completions.create(
    model="deepseek-v4.1-flash-1017-uncensored",
    messages=[{"role": "user", "content": "Hello"}],
)
print(r.choices[0].message.content)

More

Longer write-up.

Background on these builds and how they were made: DeepSeek V4.1 Flash write-up. Every ranked model: uncensored LLM leaderboard.

FAQ

Questions people ask.

Where can I run an uncensored DeepSeek V4.1 model through an API?

You can run one on OpenUncensored, which serves 8 uncensored DeepSeek V4.1 Flash builds through one OpenAI and Anthropic compatible API (https://api.openuncensored.com/v1), from Venice, heabsy, NanoGPT, Lebrel, Refuseless and UncensuredAI, with automatic fallback to the next host when one fails and no account to create. The best one on our tests is DeepSeek V4.1 Flash 1017 Uncensored (deepseek-v4.1-flash-1017-uncensored): ranked #1 of 34 on the OpenUncensored Score (69.8), answers 100% of 36 requests censored models refuse and runs at about 239 tok/s, from $0.33 per million input and $1.32 per million output tokens.

Which uncensored DeepSeek V4.1 Flash should I use?

DeepSeek V4.1 Flash 1017 Uncensored (deepseek-v4.1-flash-1017-uncensored) scores best on our tests; DeepSeek V4.1 Flash Uncensored is the cheapest. All are on the same key, so switching is a change of the model id.

Does it work with tool calling and coding agents?

Yes. Tools go through the standard OpenAI tools field or the Anthropic tool format, so Claude Code, Codex, OpenCode, Cline and Cursor work by changing the base URL. DeepSeek V4.1 Flash 1017 Uncensored finished 2/5 of our five multi-step agent tasks.

How much does it cost?

From $0.12 per million input and $0.45 per million output tokens (DeepSeek V4.1 Flash Uncensored). Prompt tokens a host serves from its cache are billed at its lower cached rate, which is most of an agent's input.

Do I need an account?

No. Buy a prepaid key on the credits page and use it as the API key. The key is the account, and it cannot spend more than you loaded.

Updated November 18, 2026. Scores measured November 8, 2026. Markdown version: /uncensored-deepseek-v4-1-api.md