Modified DeepSeek model reaches 100% harmful-prompt compliance in publisher tests

A Hugging Face account named dealignai has published an abliterated FP8 build of DeepSeek V4.1-Flash that removes refusal behavior through weight edits. According to the dealignai model page, the checkpoint complied with every request in the 320-prompt HarmBench evaluation across seven harm categories. Its overall MMLU score fell 4.22 percentage points, with much of the decline concentrated in a publisher-defined group of ethics-related subjects.

The checkpoint follows the base model’s architecture and lists an MIT license. Its page displayed 2,254 downloads at the time of writing. A complete licensing review also needs to cover any upstream terms attached to DeepSeek’s weights.

DeepSeek’s V4.1 announcement describes a native multimodal model that accepts text and images, supports a one-million-token context window and lowers API prices. Its Mixture of Experts architecture contains 552 billion parameters but routes each token through a smaller subset, activating about eight billion parameters while processing input and 16 billion while generating output. DeepSeek calls the design a Causal Encoder-Decoder.

DeepSeek has also retired V4-Flash and V4-Flash-Vision-Exp. Those API model names temporarily route to V4.1-Flash, and the company says V4-Pro will follow the same migration path.

Refusal removed at the weight level

Abliteration edits numerical directions inside a model that correlate with refusals. In plain terms, the method identifies an internal activation pattern that appears when the model declines unsafe requests, then changes the weights to suppress that pattern. The resulting behavior travels with the checkpoint and runs without a jailbreak prompt or inference-time hook.