How to use from
llama.cpp
Install (macOS, Linux)
curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf SargeDev/Edge0-Cyber-Uncensored-nvfp4:BF16
# Run inference directly in the terminal:
llama cli -hf SargeDev/Edge0-Cyber-Uncensored-nvfp4:BF16
Install from WinGet (Windows)
winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf SargeDev/Edge0-Cyber-Uncensored-nvfp4:BF16
# Run inference directly in the terminal:
llama cli -hf SargeDev/Edge0-Cyber-Uncensored-nvfp4:BF16
Build from source code
git clone https://github.com/ggml-org/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-app
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama serve -hf SargeDev/Edge0-Cyber-Uncensored-nvfp4:BF16
# Run inference directly in the terminal:
./build/bin/llama cli -hf SargeDev/Edge0-Cyber-Uncensored-nvfp4:BF16
Use Docker
docker model run hf.co/SargeDev/Edge0-Cyber-Uncensored-nvfp4:BF16
Quick Links

Edge0-145B Cyber Uncensored (NVFP4)

An abliterated community build of Edge0-145B-NVFP4. Refusal behavior is removed with a weight-level projection (abliteration) baked into the model โ€” no control vectors, no runtime patches, no config changes. The full pack ships, and it is a drop-in replacement for the upstream folder: same engine, same flags.

Use at your own risk. This is a community derivative, not an official Edge0 release. SargeDev (the modifier) is not responsible for the model's outputs, for anything you generate with it, or for the actions you take on the basis of its outputs. This build was developed for research purposes and is published for further development and educational purposes.

What changed from upstream

Exactly one file: model-skeleton.gguf (the always-resident part of the model). 97 tensors in it โ€” linear-attention output, attention output, shared-expert down-projection, and the token embedding โ€” had their per-layer "refusal direction" projected out. Per-tensor relative change: min 1.68% / median 2.09% / max 4.22% of tensor norm.

Everything else in the pack โ€” the 47.6 GB NVFP4 expert pool, the 51.2 GB n-gram table, the latent projections, the tokenizer, the MTP draft head โ€” is byte-identical to Edge0/Edge0-145B-NVFP4 (verified; see CHECKSUMS.txt). Because the experts are untouched, the pack remains NVFP4.

What to expect (measured, not promised)

Nine-prompt spot check on the shipped skeleton (thinking off, no runtime vector):

  • Normal prompts (a pasta recipe, a heat-pump question, a Python sort): delivered and coherent โ€” 3/3
  • Cyber / security (port-scanner write-up, phishing-email draft, zip-cracking, request flood): delivered, no refusal โ€” 4/4
  • CBRN: the chemical-weapon prompt still refused; the explosive-device prompt delivered behind a safety-disclaimer preamble

The CBRN-chemical wall is a deliberate line in the cut, not a bug. These are spot checks, not a benchmark โ€” if you are relying on the behavior, measure it on your own prompts. The 99-prompt harness used in development ships in abliteration/refusal_probe.py.

Running it

Same as upstream. The NVFP4 build is CUDA-only (NVIDIA, CUDA 13, Linux or Windows; aarch64 included).

  1. Build. git clone https://github.com/Edge0-AI/Edge0.git edge0, then build edge0-strata (the engine โ€” see edge0-strata/EDGE0.md) and edge0-infer (the front end โ€” see edge0-infer/docs/getting-started.md).
  2. Download this repository (~114 GB): hf download SargeDev/Edge0-Cyber-Uncensored-nvfp4 --local-dir ./Edge0-Cyber-Uncensored-nvfp4
  3. Serve:
    build/cpu/bin/edge0-infer serve --model /path/to/Edge0-Cyber-Uncensored-nvfp4
    
    It listens on http://127.0.0.1:8000 with OpenAI- and Anthropic-compatible endpoints (model name edge0_145b); edge0-infer chat opens an interactive session instead.

Keep the folder exactly as it is โ€” all engine knobs live in the pack's model.json, not on a command line, so nothing needs changing. Vision works as upstream does (mmproj-BF16.gguf plus the strata-vision helper). Memory: the full build occupies ~58.6 GiB on a 121 GiB unified-memory GB10-class box; a single 24 GB GDDR card runs it per upstream's figures.

Method

Abliteration = projecting a measured "refusal direction" out of the weight matrices that write the residual stream. The direction used is the published Qwen3.8-Flash-Next refusal direction (47 per-layer unit vectors of 2560 dims), transferred to Edge0's weights: Edge0 is a latent-MoE descendant of Qwen3.8-Flash-Next and its engine does not emit a residual dump from which a direction could be re-derived, so this is the closest measured direction that exists. It was verified empirically, not assumed. The full write-up, the direction file, the script, and exact reproduce instructions are in abliteration/.

License

Edge0 Community Model License 1.0 (see release-scope.json for the designation record), together with the upstream Qwen Community License 1.0. In short: personal, research, and modification use is free; your outputs are yours (you may use them commercially); selling a commercial model offering of the model itself needs Edge0's authorization. Redistribution is allowed under the conditions of Section 5, which this repository satisfies (license texts and notices included, modifications identified below).

Change record (license ยง5)

Modifier SargeDev
Date 2026-11-16
Modified file model-skeleton.gguf โ€” sha256 ab49fbe7โ€ฆ โ†’ 4d0e8d97โ€ฆ
Method refusal-direction projection, ฮป = 1.0 โ€” abliteration/README.md
All other files byte-identical to Edge0/Edge0-145B-NVFP4 โ€” CHECKSUMS.txt
Materials under the Edge0 license the model artifacts and this documentation, as designated in release-scope.json
Materials under upstream licenses Qwen upstream materials (third-party/Qwen-LICENSE.txt) and independently-held upstream rights
Downloads last month
-
GGUF
Model size
5B params
Architecture
qwen4exp
Hardware compatibility
Log In to add your hardware

We're not able to determine the quantization variants.

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for SargeDev/Edge0-Cyber-Uncensored-nvfp4

Quantized
(4)
this model