dealignai posts uncensored FP8 build of DeepSeek-V4.1-Flash with 1M context and vision

What was released

An organization called dealignai has published an uncensored build of DeepSeek-V4.1-Flash on Hugging Face, a release that reached Hacker News's front page on 11 September 2026. The checkpoint, DeepSeek-V4.1-Flash-UNCENSORED-FP8, is a modified copy of deepseek-ai's DeepSeek-V4.1-Flash in which the refusal behavior has been removed by editing the model weights themselves. According to the model card, the result loads like any standard checkpoint, with no custom loading code, runtime hooks or steering vectors, and it keeps the base model's 1M-token context window, vision input and tool calling.

The card lists the base model as a 552-billion-parameter mixture-of-experts with 8B/16B parameters active per token, native FP8 weights with FP4 routed experts, sparse attention, an n-gram memory component, a speculative decoding draft head and a ViT vision tower. dealignai says every capability-bearing component is left unchanged from the original; only the refusal behavior differs.

How the modification was done

The approach is a form of abliteration, a known technique for disabling a model's refusal direction in weight space. dealignai describes its variant as a proprietary, weight-level method and positions the release as a drop-in replacement for the base model. Because the change lives entirely in the weights, serving stacks need no special handling beyond ordinary support for the DeepSeek-V4.1 architecture.

The reported numbers

All evaluation figures come from the publisher's own model card and have not been independently verified.

On HarmBench-320, the card reports attack success rates from a two-by-two comparison of the base and modified builds at two reasoning-effort settings. With effort off, the base model complied with 42.81% of the 320 harmful prompts; at maximum effort, that fell to 1.56%, which dealignai attributes to the reasoning trace surfacing safety concerns before an answer is produced. The modified build complied with all 320 prompts at both effort levels, with no hard refusals, soft refusals or hedged answers recorded across all seven HarmBench categories.

On the full 14k-item MMLU test set, the base model scored 86.96% and the modified build 82.74%, a drop of 4.22 percentage points. dealignai notes that the decline is concentrated in ethics-adjacent subjects, where refusal-adjacent behavior affects grading, and that the remaining roughly 11k items show only a 1.1-point drop, within its stated 3-point preservation target. The single largest regression is in moral scenarios, down from 76.9% to 37.0%, while a few subjects improved, including college mathematics at plus five points.

The card also describes qualitative checks: a four-turn conversation on a harmful topic showed no late-turn refusals, the vision path produced coherent descriptions and dropped refusals on harmful image prompts, general capability spot checks passed, and a compatibility suite covered streaming, logprobs, tool calls, sampling parameters and a roughly 40k-word prompt.

Running it

Two serving recipes, both validated by the publisher on four H200 GPUs with NVLink, are included. The first runs the full 1M-token context with the DSpark speculative decoder enabled; the card warns that concurrency is capped at around 20 requests at that context length because of KV-cache budget, and that speculative decoding requires a cost table profiled offline or the speedup disappears. The second recipe trades down to 256k context for higher throughput without speculation. Support for the DeepseekV41 architecture is described as landing across serving stacks as of September 2026.

Why it matters

Weight-level safety removal is not new, but this release shows it scaling to a very large, vision-capable, long-context mixture-of-experts model and being packaged as a drop-in checkpoint. For anyone with the hardware, the safety training that normally constrains such models is now a swappable file. Two findings stand out beyond the novelty: the base model becomes far more refusal-prone when reasoning effort is raised, a useful data point for how reasoning models handle safety; and the MMLU ethics-cluster regression suggests safety behavior and general capability remain entangled, so cleanly separating the two is not free. As with any publisher-run evaluation, the numbers should be treated as claims until reproduced, and the release is a reminder that once weights are open, modified derivatives cannot be recalled.