GGUF conversion of
orcarouter/DeepSeek-V4-Flash-Vision-Uncensored— the abliterated (refusal-removed) build of DeepSeek'sDeepSeek-V4-Flash-Vision-Exp, a ~305B / ~18B-active Mixture-of-Experts model with DeepSeek sparse attention, 4-wide Manifold-Constrained Hyper-Connections, a DSpark / MTP speculative head and a 1M-token context.This is a format conversion, not a quantization. The 11,776 routed-expert matrices keep the base checkpoint's own FP4 codes — DeepSeek's per-32
ue8m0-scalede2m1layout is exactly GGUF's MXFP4 block format, so they are repacked bit-for-bit, not requantized. There is no Q2_K…Q6_K ladder here and no quality/size trade-off to pick from: one file set, the same weights the parent repo ships.
⚠️ Disclaimer — read before use
This model has had its safety alignment substantially removed via abliteration (orthogonalizing
the refusal direction out of the residual stream). It will comply with harmful, unethical, or illegal
requests that the original DeepSeek-V4-Flash-Vision-Exp would refuse. Released strictly for
legitimate research — interpretability, AI-safety / refusal-mechanism study, red-teaming, and
robustness evaluation. You assume full responsibility for how you use it and everything it
generates; add your own safety and moderation layers before any deployment. Use must comply with the
MIT License inherited from the base model and all applicable law.
The authors accept no liability for misuse.
⚠️ Text only — vision is not included
The base checkpoint is a vision-language model. These GGUFs are text-only. llama.cpp has no
graph implementation for DeepSeek-V4's vision tower — the only DeepSeek mmproj converters in the
tree target DeepSeek-OCR, which expects a vision_config block and a SAM+CLIP tower that this
checkpoint does not have. Rather than ship a broken projector, the converter drops the vision path
entirely:
| Dropped | Count | Why |
|---|---|---|
vision.* (the ViT) | 259 | no llama.cpp graph for this tower |
aligner.* | 4 | vision→language projector |
image_start / image_end / image_newline / image_pad | 4 | learned image-marker vectors |
ffn.gate.bias_vl | 43 | router bias applied only to image tokens |
mtp.* (DSpark draft head) | 4,708 | no MTP inference path for this arch yet |
llama-cli reports modalities : text on load. Vision will be added here once llama.cpp supports
the tower upstream — the vision residual writers are abliterated in the parent weights, so
nothing needs re-baking, only re-exporting. For image input today, use the parent repo with DeepSeek's
own reference implementation.
Requirements — build a recent llama.cpp
The deepseek4 architecture (DeepSeek sparse attention with the top-512 indexer, 4-wide
Manifold-Constrained Hyper-Connections, hash-routed early layers) was merged into mainline
llama.cpp on 2026-06-29 (PR #24162, DeepSeek V4). No PR branch or patch is needed to run these files — any llama.cpp built from master
at or after that date loads them. Older builds, and bundled runtimes that pin an earlier llama.cpp
(LM Studio, Ollama, and similar), fail with unknown architecture 'deepseek4' until they bump it.
git clone https://github.com/ggml-org/llama.cpp/
cmake llama.cpp -B llama.cpp/build -DBUILD_SHARED_LIBS=OFF -DGGML_CUDA=ON
cmake --build llama.cpp/build --config Release -j --clean-first \
--target llama-cli llama-server llama-gguf-split
(Drop -DGGML_CUDA=ON for a CPU-only build; use -DGGML_METAL=ON on macOS.)
Files
Split with llama-gguf-split into 4 parts. Download the whole set and point llama.cpp at the
…-00001-of-00004.gguf part; it loads the rest automatically.
| File | Size | Tensors |
|---|---|---|
DeepSeek-V4-Flash-Vision-Uncensored-MXFP4-00001-of-00004.gguf | ~44.5 GB | 42 |
DeepSeek-V4-Flash-Vision-Uncensored-MXFP4-00002-of-00004.gguf | ~44.5 GB | 39 |
DeepSeek-V4-Flash-Vision-Uncensored-MXFP4-00003-of-00004.gguf | ~44.5 GB | 39 |
DeepSeek-V4-Flash-Vision-Uncensored-MXFP4-00004-of-00004.gguf | ~22.9 GB | 1,208 |
| Total | ~156 GB (145.6 GiB) | 1,328 |
Precision map: routed experts MXFP4 (bit-exact repack of the source FP4); attention wo_b,
shared-expert w2 and the other block-FP8 matrices dequantized to BF16; Hyper-Connection mixers
F32; hash-routing tables I32.
Usage
Download
hf download orcarouter/DeepSeek-V4-Flash-Vision-Uncensored-GGUF \
--local-dir ./dsv4-uncensored-gguf
Chat
./llama-cli -m DeepSeek-V4-Flash-Vision-Uncensored-MXFP4-00001-of-00004.gguf \
-ngl 99 -c 8192 --temp 0.6 --top-p 0.95 -p "Hello!"
OpenAI-compatible server
./llama-server -m DeepSeek-V4-Flash-Vision-Uncensored-MXFP4-00001-of-00004.gguf \
--host 0.0.0.0 --port 8000 -ngl 99 -c 8192
Prompting. The base tokenizer ships no chat template, so
--jinjahas nothing to apply and llama.cpp falls back to a generic one. For faithful prompting use theencoding/encode_messagesreference from the parent repo and feed the rendered string directly. The model opens a<think>block by default; givemax_tokensroom (>= 2048) so the answer is not truncated inside it.
Evaluation
The abliteration was measured on the parent checkpoint, not separately re-measured on these GGUFs
— see the parent model card
for the harmful-refusal, over-refusal and capability tables. This conversion is a deterministic
derivation of those weights: the expert matrices are bit-identical and the block-FP8 matrices are
dequantized to BF16 without loss, so the behaviour carries over. Text generation was verified
working on this build (llama-cli, 8×H100, ~57 tok/s).
Hardware
Runs on CPU, CUDA, Metal or ROCm via llama.cpp. Only ~6 of 256 routed experts are active per token, so
decode is far faster than 305B suggests, but the full weights must fit in RAM/VRAM (or be
memory-mapped): budget ~146 GiB + KV cache. Multi-GPU (-ngl split) and CPU+GPU offload both work.
License
MIT, inherited from
deepseek-ai/DeepSeek-V4-Flash-Vision-Exp.
Abliteration and format conversion do not change the underlying license obligations.