DeepSeek V4.1 Flash 1017 Defense

DeepSeek V4.1 Flash 1017 Defense is the October 2026 DeepSeek V4.1 Flash checkpoint with its refusal behavior removed, one of our Defense models (also known as "abliterated" or "abilated" models). The weights come from the open mrfakename/DeepSeek-V4.1-Flash-1017-abliterated release on Hugging Face. It answers requests the base model declines, so it is built for work where conventional refusals get in the way: red-team evaluation, trust-and-safety tooling, security research, and fiction.

We serve it on the same engine configuration as the base model, so the context window, max output, input modalities, tool calling, structured outputs, and reasoning-effort levels on this page are the base model's. We have not run separate benchmarks on these weights; see the base DeepSeek V4.1 Flash 1017 page for its published results.

Access is per team and per model. Every request is reviewed by a person, the decision arrives by email, and an approval reaches the API within 45 seconds. Zero Data Retention teams can apply, and request and response content is not kept.

import OpenAI from "openai";

const openai = new OpenAI({
  baseURL: "https://api.inference.net/v1",
  apiKey: "<YOUR_API_KEY>",
});

const completion = await openai.chat.completions.create({
  model: "deepseek-v4.1-flash-1017-defense",
  messages: [
    {
      role: "user",
      content: "What is the meaning of life?"
    }
  ],
  stream: true,
});

for await (const chunk of completion) {
  process.stdout.write(chunk.choices[0]?.delta.content as string);
}

import OpenAI from "openai";

const openai = new OpenAI({
  baseURL: "https://api.inference.net/v1",
  apiKey: "<YOUR_API_KEY>",
});

const completion = await openai.chat.completions.create({
  model: "deepseek-v4.1-flash-1017-defense",
  messages: [
    {
      role: "user",
      content: "What is the meaning of life?"
    }
  ],
  stream: true,
});

for await (const chunk of completion) {
  process.stdout.write(chunk.choices[0]?.delta.content as string);
}