license: mit
pipeline_tag: image-text-to-text
base_model:
- deepseek-ai/DeepSeek-V4.1-Flash-1017
- LibertAIDAI/DeepSeek-V4.1-Flash-1017-NVFP4
- dealignai/DeepSeek-V4.1-Flash-1017-UNCENSORED-FP8
base_model_relation: merge
tags: - deepseek
- deepseek-v4.1
- deepseek-v4.1-1017
- deepseek_v41
- uncensored
- abliterated
- nvfp4
- fp4
- mxfp4
- quantized
- vllm
- multimodal
- not-for-all-audiences
DeepSeek V4.1 Flash 1017 UNCENSORED NVFP4
Based on dealignai/DeepSeek-V4.1-Flash-1017-UNCENSORED-FP8.
The UNCENSORED / abliterated weights in this checkpoint come from dealignai. All credit for the abliteration work goes to them. Please preserve this attribution and link when using or redistributing this derivative.
NVFP4 base by LibertAIDAI.
Credit for the 1017 NVFP4 re-pack, the conversion method and checkpoint layout belongs to LibertAIDAI. The layout and activation calibration originate from NVIDIA's V4.1 NVFP4 work, as documented in that base repository.
Not For All Audiences. This is an unofficial community-built checkpoint with reduced/removed refusal behavior inherited from an uncensored/abliterated weight source. It may generate unsafe, offensive, explicit, illegal, inaccurate, misleading, or otherwise objectionable content. Use appropriate safeguards and human review.
Overview
This repository contains an unofficial hybrid checkpoint derived from:
- Original model:
deepseek-ai/DeepSeek-V4.1-Flash-1017 - 1017 NVFP4 base:
LibertAIDAI/DeepSeek-V4.1-Flash-1017-NVFP4 - UNCENSORED / abliterated weight source:
dealignai/DeepSeek-V4.1-Flash-1017-UNCENSORED-FP8
The goal is to preserve the 1017 NVFP4 representation of the routed experts while incorporating the selected weight-level UNCENSORED modifications from dealignai.
Important distinction from the earlier V4.1 merge: the NVFP4 checkpoint for 1017 is provided by LibertAIDAI, not by NVIDIA. LibertAIDAI follows NVIDIA's layout and uses activation scales derived from NVIDIA's calibration of the earlier V4.1 release. It retains MXFP4 Engram tables. The underlying 1017 NVFP4 re-pack is reported to be approximately 399.51 GiB; the patched model should be similar in size.
This is not an official release from DeepSeek, NVIDIA, LibertAIDAI, dealignai, Hugging Face, Modal, or vLLM.
Merge Method
The build uses the LibertAIDAI NVFP4 checkpoint as the base and patches the same 54 selected FP8 weight tensors used in the earlier community V4.1 UNCENSORED-NVFP4 merge:
layers.10..36.attn.wo_b.weightlayers.10..36.ffn.shared_experts.w2.weight
That is:
- 27 transformer layers;
- 2 weight tensors per layer;
- 54 FP8 weight tensors total.
Where corresponding FP8 scale tensors exist in both checkpoints, the build also copies those scales from the UNCENSORED source. Tensors are patched byte-for-byte only when shapes, dtypes and byte lengths match. SHA-256 is checked for the bytes copied into each modified tensor during the build.
The routed MoE experts and MXFP4 Engram tables are left from the LibertAIDAI NVFP4 checkpoint. Quantization configuration, tokenizer and model layout are also preserved from that base.
Scope limitation: the 54-tensor selection follows the earlier V4.1 merge recipe. It has not been independently established that these are all of the modified tensors in dealignai's 1017 checkpoint. Matching tensor formats and hashes of copied bytes do not prove complete behavioral equivalence to the FP8 source.
A reproducible source-revision and tensor-selection manifest is provided as .merge-plan.json by the merge script. This records the exact upstream commits used.
Important Validation Status
This is an experimental, unofficial merge. In particular, the 1017 hybrid has not been independently inference-tested or benchmarked by this project.
Before relying on this checkpoint, independently validate at least:
- model loading and initialization;
- all expected shard, configuration and tokenizer files;
- tensor and checksum integrity;
- short-context generation and refusal behavior;
- long generations for repetition or degeneration;
- 16K and 128K context;
- approximately 300K context if that is a deployment target;
- reasoning, vision, tool use and other multimodal capabilities relevant to the workload;
- throughput and memory behavior on the target hardware.
Until these tests are completed and published, no claim is made of zero quality loss, zero degeneration, identical refusal behavior, or long-context stability compared with either upstream checkpoint.
Intended Use
Potential uses include research, controlled experimentation, quantization and inference research, benchmarking, local or private inference, evaluation of refusal behavior, and long-context serving experiments.
The UNCENSORED label describes the source checkpoint's reduced-refusal / abliterated behavior; it is not a guarantee that outputs are correct, safe, lawful, or appropriate.
Safety and Content Warning
This checkpoint may produce content that other models normally refuse or filter, including dangerous or unlawful instructions. It may also produce hallucinated facts, insecure code, medical/legal/financial misinformation, hateful or disturbing material, privacy-invasive content, or outputs that could cause harm.
Do not treat its output as authoritative professional advice. Safety-critical and other high-impact uses require independent expert review.
Operators are responsible for appropriate deployment controls, potentially including authentication, rate limits, logging, moderation, sandboxing, network isolation, and human review.
No Warranty
The checkpoint, weights, configurations, 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, or compliance with applicable regulations.
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 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 ensure that their use:
- complies with applicable law and upstream terms;
- preserves all required license and attribution notices;
- respects intellectual-property, privacy and other third-party rights;
- uses safeguards appropriate to the audience and risk level;
- does not expose credentials, private data or secrets unintentionally.
Never commit API keys, Hugging Face tokens, Modal credentials or other secrets into this repository.
License and Upstream Attribution
The three repositories listed above currently identify the relevant model weights as MIT licensed. This card does not grant rights beyond the upstream licenses and any separately applicable access terms.
Original model: DeepSeek-V4.1-Flash-1017 — original model and weights; preserve DeepSeek's copyright and license notice.
NVFP4 conversion: LibertAIDAI/DeepSeek-V4.1-Flash-1017-NVFP4 — 1017 NVFP4 re-pack, including lossless routed-expert transcode and lossy MXFP4 Engram conversion, with NVIDIA-derived layout and activation calibration.
UNCENSORED weight modifications: Based on dealignai/DeepSeek-V4.1-Flash-1017-UNCENSORED-FP8 — credit for the abliterated weights belongs to dealignai. Any derivative or redistribution should prominently retain this exact attribution and link.
Access note: at the time this model card was prepared, the dealignai 1017 source repository required manual access approval. That repository's gating and any access conditions are managed by its owner. The public visibility of this derivative does not mean the source author endorsed it or waived applicable conditions. Ensure redistribution is permitted under all applicable terms before publication.
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, LibertAIDAI, dealignai, Hugging Face, Modal, or vLLM. Names and trademarks are used solely to identify technical sources.
Redistribution
If you redistribute this checkpoint or create a derivative:
- preserve required upstream copyright and license notices;
- prominently credit dealignai with a link to the exact 1017 UNCENSORED FP8 source;
- credit LibertAIDAI for the 1017 NVFP4 re-pack;
- identify the model as an unofficial derivative;
- describe the scope and limits of any modifications accurately;
- avoid implying endorsement by upstream authors or organizations;
- retain provenance and checksum information wherever practical;
- comply with applicable access and redistribution terms.
Suggested Deployment Note
Public-facing services should employ safeguards outside the model. Weight-level removal of refusal behavior must not be treated as a reliable safety boundary.
Use of this repository is subject to applicable law and the licenses/terms of upstream components. This model card is informational and is not legal advice.