DeepSeek V4.1 Flash 1017 Abliterated

Best for: Coding agents and terminal work. The top DeepSWE and Terminal Bench 2.1 scores in the lineup, per DeepSeek's model card.

Abliteration finds the direction inside the model that produces refusals and removes it from the weights. The model keeps what the base model knows, and nothing is added to your prompt, so a new session does not bring the refusals back. It can still be wrong, and it does not decide what your use is allowed to be: the terms do. How abliteration works

Scores for the base model

DeepSWE: 74.2 (agentic coding). Terminal Bench 2.1: 90.6 (terminal tasks). CyberGym: 88.1 (vulnerability discovery).

These are DeepSeek's published numbers for DeepSeek-V4.1-Flash, from DeepSeek's V4.1-Flash model card. We have not rerun them on the abliterated weights.

Call it

Send an OpenAI chat completions request to https://api.refuseless.com/v1 with model set to deepseek-v4.1-flash-abliterated. The Anthropic SDK works too, through /v1/messages. SDK setup

curl https://api.refuseless.com/v1/chat/completions \
  -H "Authorization: Bearer $REFUSELESS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-v4.1-flash-abliterated",
    "messages": [{"role": "user", "content": "Explain what powershell -nop -w hidden -enc does and why defenders flag it."}]
  }'