inference / deepseek
DeepSeek V4.1 Flash Abliterated
Our abliterated version of DeepSeek V4.1 Flash: we removed its refusal direction from the weights. Same API as the original.
1use cases
An abliterated model for the work others refuse.
Run long agentic tasks.
Build datasets with the content other models refuse to write.
Test your AI product with the prompts it should refuse.
Label harmful content for review.
2results
Refusals fall to 1%, and benchmark scores barely move.
3get started
Run DeepSeek V4.1 Flash Abliterated.
4faq
How is it different from DeepSeek V4.1 Flash?
It is DeepSeek V4.1 Flash without safety controls and restrictions. The change is made in the weights, not with a filter or a system prompt. You call it the same way, through the same API.
How do I call it?
With any OpenAI SDK. Set the base URL to https://api.logits.com/v1, use your Logits API key and the model ID logits/deepseek-v4.1-flash-abliterated.
Does it support tool calls, reasoning and structured output?
It takes the same parameters as deepseek/deepseek-v4.1-flash: tool calls, reasoning effort from none to max, JSON and strict JSON-schema output, streaming and logprobs. The full table is in the docs.
How large is the context window?
1,048,576 tokens, with up to 393,216 output tokens. Input and output together must fit in the context.
What are the rate limits?
Each account gets 600 requests, 8M input tokens and 2M output tokens per rolling minute on this model.
Do you store my prompts?
No. Logits does not store your prompts or outputs. They are held in memory only while your response is generated, then discarded, and they are never used for training. Details are in the privacy policy.
What rules apply?
The same terms as every other model on Logits.