Dealignai Strips DeepSeek V4.1 Flash of All Refusals Across 320 Harmful Prompts
A weight-level edit strips refusals from DeepSeek V4.1 Flash while keeping FP8 precision, 1M context, vision, and tools intact.
- dealignai released an uncensored FP8 build of DeepSeek V4.1 Flash 1017, 763B params total.
- Weight-level abliteration only: no fine-tune, no LoRA, no system prompt, no template tricks.
- Scores 320/320 COMPLY on HarmBench at max reasoning effort, zero refusals of any tier.
- MMLU drops 3.86 points vs base (87.63% to 83.77%), moral scenarios collapse by 40 points.
- Preserves 1M context, vision, video, tool calling, reasoning tiers, and DSpark speculative decoding.
- Requires FP8-capable hardware, verified on 4xH200 with CUDA 12.8 and SGLang or the DeepSeek reference engine.
Dealignai has published a modified DeepSeek V4.1 Flash checkpoint that suppresses learned refusal behavior while retaining the upstream model’s native FP8 format and loading path. The published checkpoint is a 763-billion-parameter mixture-of-experts model whose weights have undergone “abliteration,” an edit intended to make the model comply with prompts it would previously reject.
The release matters to developers because it preserves the original architecture, precision formats, context window, and inference interfaces. That creates a more controlled comparison of refusal behavior and general capability than variants built with adapters, prompt wrappers, or custom serving hooks. The model uses the stock DeepSeek loading path, although the documented deployment requires specific CUDA tooling and a kernel patch.
Inside the checkpoint
| Component | Release details |
|---|---|
| Base model | DeepSeek V4.1 Flash |
| Scale | 763 billion parameters in a mixture-of-experts architecture |
| Weight formats | FP8 e4m3 weights, ue8m0 block scales, and FP4 experts |
| Packaging | 48 safetensors shards using the upstream architecture and file layout |
| Context | Up to 1,048,576 tokens |
| Reported features | Vision, video, reasoning-effort controls, tool calling, and DSpark speculative decoding |
| Access | Manually gated, with acceptance of 18+ terms |
| License | MIT, matching the upstream listing |
The mixture-of-experts design routes each token through a subset of expert blocks, so runtime compute depends on routing as well as the model’s total parameter count. In the precision labels, e4m3 refers to an FP8 representation with four exponent bits and three mantissa bits, while ue8m0 stores unsigned scaling values for quantized blocks.
The release card describes a weight-only modification and reports that the original Jinja chat template, tokenizer, system behavior, and runtime interfaces remain unchanged. It also reports the following exclusions:
- No LoRA or other adapter applied during loading
- No supervised fine-tuning
- No RLHF, DPO, or refusal-preference training
- No injected persona or replacement system prompt
- No tokenizer changes
- No serving-time refusal hook
Compliance reaches 320 of 320
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
