license: mit
pipeline_tag: text-generation
base_model: Jiunsong/SuperDeepseek-V4-Flash-abliterated
base_model_relation: quantized
language: [en]
tags: [gguf, quantized, deepseek, deepseek-v4, deepseek-v4-flash, mixture-of-experts, moe, mxfp4, fp4, ds4, dwarfstar, apple-silicon, metal, abliterated]

SuperDeepseek V4 Flash Abliterated — DS4 MXFP4

A DS4-targeted MXFP4 GGUF conversion of Jiunsong/SuperDeepseek-V4-Flash-abliterated, itself derived from deepseek-ai/DeepSeek-V4-Flash-0731.

This repository contains a single approximately 156 GB GGUF built specifically for the DwarfStar / DS4 inference engine.

Model file

Exact output size:

155,976,458,848 bytes

What this is

The source model is Jiunsong's SuperDeepseek V4 Flash abliterated checkpoint.

SuperDeepseek modifies selected attention/output components of DeepSeek V4 Flash to substantially reduce unnecessary refusals while retaining the original routed MoE experts.

This conversion repackages that checkpoint into the specialized GGUF layout expected by DS4.

The routed experts use DS4's native MXFP4 expert path, preserving the source checkpoint's released FP4 expert representation rather than replacing the experts with an IQ2/Q2 quantization.

Conversion provenance

This GGUF was generated from the following exact inputs:

Source repository:
Jiunsong/SuperDeepseek-V4-Flash-abliterated

Source revision:
93f0b5e3d1087d1e2ad12958fb89a394defe12e1

DS4 template repository:
antirez/deepseek-v4-gguf

Template revision:
e7f04037032990db0346398d249baf9fb9df1ccc

Template file:
DeepSeek-V4-Flash-MXFP4Experts-F16HC-F16Compressor-F16Indexer-Q8Attn-Q8Shared-Q8Out-chat-v2-mxfp4-0731.gguf

DS4 converter commit:
84cc882352757baf628a1776badf7cc54d584e28

Output:
SuperDeepseek-V4-Flash-abliterated-DS4-MXFP4-0731.gguf

Output size:
155976458848 bytes

The conversion was performed using DS4's native deepseek4-quantize tooling.

CONVERSION_INFO.txt is included in this repository as a machine-readable record of the conversion provenance.

Intended runtime: DS4

This model was produced specifically for DS4 / DwarfStar.

Clone and build DS4 on macOS:

git clone https://github.com/antirez/ds4.git
cd ds4
make

Then run this GGUF by specifying it explicitly:

./ds4 \
  -m /path/to/SuperDeepseek-V4-Flash-abliterated-DS4-MXFP4-0731.gguf \
  -p "Explain the architecture of DeepSeek V4 Flash."

Running with SSD streaming

The GGUF is much larger than the unified memory available on many Macs.

DS4 supports SSD streaming for this situation:

./ds4 \
  -m /path/to/SuperDeepseek-V4-Flash-abliterated-DS4-MXFP4-0731.gguf \
  --ssd-streaming \
  -p "Explain the architecture of DeepSeek V4 Flash."

With SSD streaming, DS4 keeps non-routed model weights resident while dynamically loading routed MoE experts from the GGUF into an in-memory expert cache.

The automatic cache budget is the recommended starting point.

Available memory, context length, SSD performance, and memory pressure will affect whether a particular system can run the model comfortably and how quickly it generates.

Important compatibility note

This GGUF targets DS4.

Hugging Face may automatically display usage instructions for llama.cpp, Ollama, LM Studio, or other GGUF-compatible applications because the uploaded artifact is recognized as a GGUF model.

Those automatically generated instructions should not be interpreted as confirmation that this particular DS4-oriented tensor layout has been validated with those runtimes.

For the intended and tested loading path, use DS4.

About the source model

Jiunsong/SuperDeepseek-V4-Flash-abliterated is based on:

deepseek-ai/DeepSeek-V4-Flash-0731

The source release reports that the SuperDeepseek modification targets the 43 backbone and 3 MTP attn.wo_b weight/scale pairs plus a bounded output-head recovery, while leaving the routed experts untouched.

For details on the abliterated model, its methodology, benchmark results, long-context testing, and limitations, refer to the original Jiunsong model card.

Credits

SuperDeepseek

Original modified checkpoint:

Jiunsong/SuperDeepseek-V4-Flash-abliterated

Created by Jiunsong.

DeepSeek V4 Flash

Base model:

deepseek-ai/DeepSeek-V4-Flash-0731

Created by DeepSeek.

DS4 / DwarfStar

GGUF format, MXFP4 template, conversion tooling, and intended inference runtime:

antirez/ds4
antirez/deepseek-v4-gguf

Created and maintained by Salvatore Sanfilippo (antirez) and contributors.

This conversion

DS4 MXFP4 conversion and repository:

KarlKinda/SuperDeepseek-V4-Flash-abliterated-ds4-MXFP4

Responsible use

"Abliterated" refers to modification of model refusal behavior. It does not guarantee correctness, reliability, safety, or suitability for any particular use.

Users remain responsible for evaluating outputs and applying appropriate safeguards for their deployment.

License

MIT, following the upstream DeepSeek V4 Flash and SuperDeepseek releases.

The original model authors retain their respective copyrights and attribution.