Dealignai posts abliterated DeepSeek V4.1 Flash weights on Hugging Face

The dealignai research team published an abliterated build of DeepSeek-V4.1-Flash to Hugging Face, filed as DeepSeek-V4.1-Flash-UNCENSORED-FP8. The repository reached the Hacker News front page on Sept. 11, 2026, where it drew 25 points and 2 comments. The model card summarizes the release in a single line: abliterated, no guardrails, native FP8, a 1-million-token context window, and vision plus tool support.

The checkpoint derives from DeepSeek-V4.1-Flash, the vision-language Mixture-of-Experts model DeepSeek published as open weights. The Hugging Face listing classifies the upload as an image-text-to-text transformer using the deepseek_v41 architecture, stored in safetensors and quantized to fp8 8-bit precision. Its tags include abliterated, uncensored, crack, multimodal, Mixture of Experts and text-generation. Abliteration edits a model's weights to remove the internal directions that drive refusal behavior, so the checkpoint answers prompts the base model declines.

The model card makes a specific technical case for how this build differs from earlier uncensored repackages. Dealignai describes the method as proprietary weight-level abliteration: the refusal circuitry is surgically removed, and no custom model.py, runtime hooks or steering vectors are involved. The result, the card states, is a standard checkpoint that loads exactly like the base model. Every capability-critical component is preserved byte-identical to the base, according to the card, which names the routed experts, Engram memory, CSA2 sparse attention, DSpark draft head, vision tower, router gates, norms and embeddings.

The byte-identical claim is the release's central proposition. The card is written against the objection that uncensored builds trade capability for compliance, and its emphasis on leaving the experts, sparse attention and draft head untouched is the answer it offers. No benchmark numbers appear in the material attached to the repository listing, so the claim rests on the team's own description rather than on published evaluation results.

The upload carries an MIT license, the terms DeepSeek applies to the base model, which permits redistribution and modification. That license is why a third party can publish a safeguard-stripped copy without DeepSeek's involvement. Dealignai's repository is the latest entry in the practice of repacking open-weight releases for unrestricted use, and it inherits the base model's architecture while removing the behavior layer the original ships with.

The Hacker News submission's 25 points and 2 comments indicate limited early attention. The comments themselves are not visible in the retrieved material, so it is not yet possible to say how readers responded to the technical claims. Dealignai's own benchmarks, if any, are not part of the listing excerpt.

Open questions remain about method and verification. The card does not state whether the checkpoint was produced by abliteration alone or by additional fine-tuning, and no independent lab has reported testing the uncensored build. The practical question for users is whether the release retains the vision, tool-calling and long-context behavior of the base model, which the card asserts and which third-party evaluation would have to confirm.