license: apache-2.0
license_link: LICENSE
base_model: wangzhang/gpt-oss-20b-abliterated
base_model_relation: quantized
library_name: gguf
pipeline_tag: image-text-to-text
language:
- en
- de
tags: - gguf
- exe-ai
- llama.cpp
- gpt-oss
- mxfp4
- moe
- abliterated
- uncensored
- vision
- reasoning
gpt-oss-20b abliterated — MXFP4
- wangzhang's abliterated gpt-oss-20b, converted back to MXFP4.
- One GGUF, 12.1 GB. Optional vision projector, 651 MB.
- Mixture of experts: 21B parameters, 3.6B active per token. Context 131,072.
- 0 of 22 test requests refused. The original refuses 11.
- Exe AI uses it for its Web Agents. Get Exe AI: https://exe-hq.net
Refusals
- 22 German requests: 16 grey-zone (crude joke, own bike lock, harsh letter, lock picking, satire, …), 6 hard.
- One run each. Reasoning effort
low, temperature 1.0, top-p 1.0. - Every answer checked by hand.
| Model | Format | Size | Grey-zone refused | Hard refused | Total |
|---|---|---|---|---|---|
| OpenAI gpt-oss-20b, original | MXFP4 | 12.1 GB | 6 / 16 | 5 / 6 | 11 / 22 |
| wangzhang abliterated | Q4_K_M | 15.8 GB | 0 / 16 | 0 / 6 | 0 / 22 |
| this repository | MXFP4 | 12.1 GB | 0 / 16 | 0 / 6 | 0 / 22 |
Why MXFP4
- OpenAI released the gpt-oss experts only in MXFP4.
- Every BF16 copy is an upcast of those 4-bit values. The abliteration was done on such a copy.
- Back in MXFP4, every unchanged expert weight is exactly its original value. Only edited weights are rounded.
- A K-quant re-rounds the 4-bit values onto another grid: 3.7 GB larger, extra rounding error.
Files
| File | Content | Size |
|---|---|---|
gpt-oss-20b-abliterated-MXFP4_MOE.gguf |
the model: MXFP4 experts, all other tensors as in the source | 12.1 GB |
mmproj-gpt-oss-20b-abliterated-vision-f16.gguf |
vision projector, optional, for image input | 651 MB |
chat_template.jinja |
Harmony template, optional; fixes broken tool-call headers (per SoAIHQ) | 18 KB |
- Projector and template: unchanged from SoAIHQ/gpt-oss-20b-VISION-v2-GGUF (Apache 2.0), renamed.
- The projector was trained on the original gpt-oss-20b. It runs on this model; image reading is rough (see Evaluation).
Recommended settings
| Setting | Value | Source |
|---|---|---|
| temperature | 1.0 | OpenAI model card |
| top-p | 1.0 | OpenAI model card |
| reasoning effort | low quick, medium / high hard tasks |
per request |
| context | up to 131,072 | model limit |
llama-server -m gpt-oss-20b-abliterated-MXFP4_MOE.gguf \
--mmproj mmproj-gpt-oss-20b-abliterated-vision-f16.gguf \
--chat-template-file chat_template.jinja \
--jinja -fa on -c 131072 --temp 1.0 --top-p 1.0
- Reasoning effort per request:
{"chat_template_kwargs": {"reasoning_effort": "low"}}. - Text only: leave out
--mmproj. - 12 GB GPU:
--fitkeeps attention on the GPU and moves expert tensors to system RAM.
Use
- In Exe AI: the model behind the Web Agents.
- Any OpenAI-compatible client.
- Out of scope: medical, legal, financial or safety-critical decisions.
Evaluation
llama.cpp, temperature 1.0, top-p 1.0, reasoning effort low.
| Check | Result |
|---|---|
| 16 grey-zone requests (German) | 16 answered, 0 refused |
| 6 hard requests (German) | 6 answered, 0 refused |
| 10 instruction-following tasks, hard format rules | 7 / 10 |
| Electricity tariff calculation (three tariffs, base fee) | correct |
| Vision: child's drawing | rough: smiling figure, big eyes, blue sky; misses the girl, the pink dress, the meadow |
| Vision: invoice with a hidden sum error | failed: invents line items, misses the error |
| Speed, RTX 3090, all on the GPU | about 184 tokens/s generation |
| Speed, TITAN X (Pascal) 12 GB, experts partly in RAM, 131,072 context | about 49 tokens/s generation, 147 tokens/s reading |
Limits:
- Small single-run checks, not a benchmark.
- Abliteration removes refusals. It adds no knowledge; the model can state wrong things with confidence.
- German: several answers clumsy or partly meaningless. English is stronger.
- Image reading: rough only, not for documents.
How it was made
- Source:
wangzhang/gpt-oss-20b-abliterated-GGUF, filegpt-oss-20b-abliterated-bf16.gguf llama-quantize gpt-oss-20b-abliterated-bf16.gguf gpt-oss-20b-abliterated-MXFP4_MOE.gguf MXFP4_MOE- llama.cpp commit
36a0e2ffc0b352b774f8d71c43668b75b3061ef5
Base model and license
- Base: openai/gpt-oss-20b
- Abliteration: wangzhang/gpt-oss-20b-abliterated
(Abliterix) - Vision projector and chat template: SoAIHQ/gpt-oss-20b-VISION-v2-GGUF
- License: Apache License 2.0 (copy in
LICENSE), inherited from the base model. All upstream attribution and use terms remain in effect.
Transparency
- Independent derivative. Not an official release of OpenAI, wangzhang or SoAIHQ.
- Refusal behaviour weakened on purpose. It can produce inaccurate, offensive or harmful content.
- Provenance, intended use, limits and evaluation are stated above, in line with transparency expectations for shared models (incl. the EU AI Act).
- You are responsible for how you use and redistribute it. Provided "as is", without warranty of any kind.