license: mit
pipeline_tag: image-text-to-text
base_model:
- nvidia/DeepSeek-V4.1-Flash-NVFP4
- dealignai/DeepSeek-V4.1-Flash-UNCENSORED-FP8
base_model_relation: merge
tags: - deepseek
- deepseek-v4.1
- deepseek_v41
- uncensored
- abliterated
- nvfp4
- fp4
- quantized
- vllm
- not-for-all-audiences
DeepSeek V4.1 Flash UNCENSORED NVFP4
Not For All Audiences. This is an unofficial community-built checkpoint with reduced/removed refusal behavior inherited from an uncensored/abliterated weight source. It can generate unsafe, offensive, explicit, illegal, inaccurate, misleading, or otherwise objectionable content. Use appropriate access controls, moderation, sandboxing, and human review for your use case.
Overview
This repository contains an unofficial hybrid checkpoint derived from:
- Original model:
deepseek-ai/DeepSeek-V4.1-Flash - NVFP4 base:
nvidia/DeepSeek-V4.1-Flash-NVFP4 - UNCENSORED / abliterated weight source:
dealignai/DeepSeek-V4.1-Flash-UNCENSORED-FP8
The goal is to preserve NVIDIA's Blackwell-oriented NVFP4 representation for the routed experts while incorporating the weight-level UNCENSORED modifications from the dealignai checkpoint.
This repository is not an official release from DeepSeek, NVIDIA, dealignai, Hugging Face, Modal, or vLLM.
Merge Method
The build process uses the NVIDIA NVFP4 checkpoint as the base and replaces only the 54 weight tensors identified as modified in the UNCENSORED checkpoint:
layers.10..36.attn.wo_b.weightlayers.10..36.ffn.shared_experts.w2.weight
That is:
- 27 transformer layers
- 2 replaced tensors per layer
- 54 replaced tensors total
The routed MoE experts are not replaced by this merge and remain from the NVIDIA NVFP4 checkpoint.
The NVIDIA quantization configuration excludes the attention and shared-expert paths above from NVFP4 quantization, which is why the merge is performed directly on those source-precision weights rather than requantizing the entire model.
A build/patch manifest and checksums should be kept with the repository whenever possible so that users can verify provenance and integrity.
Important Validation Status
This is an unofficial merge. Do not assume that matching tensor shapes or successful loading proves behavioral equivalence or long-context stability.
Before relying on this checkpoint, independently validate at least:
- model loading and initialization;
- tensor/checksum integrity;
- ordinary short-context generation;
- long generations for repetition or degeneration;
- 16K context;
- 128K context;
- approximately 300K context if that is your deployment target;
- reasoning and tool-use behavior relevant to your workload;
- throughput and memory behavior on your target hardware.
Until those tests have been completed and documented, no claim is made that this merge has zero quality loss, zero degeneration, or identical behavior to either upstream checkpoint.
Intended Use
Potential uses include:
- research;
- controlled experimentation;
- quantization/inference research;
- benchmarking;
- local or private inference;
- evaluation of refusal behavior;
- long-context serving experiments.
The "UNCENSORED" label describes the source checkpoint's reduced-refusal / abliterated behavior. It does not mean that outputs are correct, safe, lawful, or appropriate.
Safety and Content Warning
This checkpoint may produce content that other models would normally refuse or filter, including potentially dangerous or unlawful instructions.
It may also produce:
- hallucinated facts or citations;
- insecure or destructive code;
- medical, legal, financial, or technical misinformation;
- hateful, harassing, sexual, graphic, or otherwise disturbing material;
- privacy-invasive content;
- instructions that could cause physical, digital, financial, or other harm.
Do not treat model output as authoritative professional advice. High-impact or safety-critical uses require independent expert review and appropriate safeguards.
Operators are responsible for deciding what controls are appropriate for their deployment, including authentication, rate limits, logging, moderation, sandboxing, network isolation, and human review.
No Warranty
The checkpoint, weights, configuration, scripts, documentation, and related materials are provided "AS IS", without warranties of any kind, express or implied, to the maximum extent permitted by applicable law.
No representation or guarantee is made regarding:
- correctness;
- accuracy;
- safety;
- reliability;
- security;
- fitness for a particular purpose;
- uninterrupted operation;
- absence of hallucination;
- absence of degeneration;
- long-context stability;
- compatibility with a particular inference stack;
- compliance with laws or regulations applicable to a particular deployment.
Use and deployment are at your own risk.
Limitation of Liability
To the maximum extent permitted by applicable law, the maintainer and contributors disclaim liability for claims, losses, damages, costs, or consequences arising from the use, misuse, modification, redistribution, hosting, or inability to use this repository or its contents.
Nothing in this notice excludes or limits liability that cannot lawfully be excluded or limited.
This section is a general project disclaimer, not legal advice.
User and Operator Responsibilities
Downloading or using this checkpoint does not transfer responsibility for a deployment to the maintainer.
Users and operators should independently ensure that their use:
- complies with applicable law;
- complies with the licenses and terms governing upstream materials;
- respects intellectual-property, privacy, contractual, and other third-party rights;
- uses appropriate safeguards for the intended audience and risk level;
- does not expose credentials, private data, or secrets to the model unintentionally.
Do not commit API keys, Hugging Face tokens, Modal credentials, or other secrets to this repository.
License and Upstream Attribution
The Hugging Face repositories listed above currently identify the relevant model weights as MIT licensed. This repository does not grant any rights beyond those provided by the applicable upstream licenses.
Preserve all copyright, license, and attribution notices required by the upstream projects.
In particular, consult the current upstream repositories before redistribution or commercial deployment:
Upstream licenses and terms control where applicable. This README does not replace them.
Non-Affiliation
This project is independently assembled by a community maintainer.
It is not affiliated with, endorsed by, sponsored by, or officially supported by DeepSeek, NVIDIA, dealignai, Hugging Face, Modal, or vLLM.
Names and trademarks are used only to identify upstream projects and technical provenance.
Redistribution
If you redistribute this checkpoint or create a derivative:
- preserve all required upstream license and copyright notices;
- identify it clearly as an unofficial derivative;
- accurately describe any modifications you make;
- avoid implying endorsement by upstream authors or organizations;
- preserve provenance information and checksums where practical.
Suggested Deployment Note
For public-facing services, consider adding safeguards outside the model itself. Weight-level removal of refusal behavior means the model should not be assumed to provide a reliable safety boundary.
Use of this repository is subject to applicable law and the licenses/terms of the upstream components. This model card is informational and is not legal advice.