Abliterated Qwen3.8-27B: Compare 4 Editions and Choose

Abliterated Qwen3.8-27B models share a base name, but their contents depend on who modified them. Each creator uses different tools, edits different layers, and publishes different numbers, so the editions behave like separate models. Alibaba released Qwen3.8-27B on August 14, 2026. The first abliterated edition appeared the next day, and a fuller edition from a model-editing workshop followed within weeks.

This article compares four editions so you can decide which to use: huihui-ai, orcarouter, trohrbaugh's Heretic-ARA, and slevinw's RVN. In short, huihui-ai limits its edit to layers 18–51 to protect performance, orcarouter was first to publish and offers 12 quantization levels, and RVN reports the lowest KL divergence at 0.0085.

---

What is the base Qwen3.8-27B model?#

Qwen3.8-27B is an open-weight dense model that Alibaba released on August 14, 2026, under the Apache 2.0 license. It accepts image and video input, supports a 262K-token context, and uses a 64-layer hybrid architecture. Its strong baseline performance made it an immediate target for community modification.

Alibaba published it on Hugging Face and ModelScope. "Open weights" means the model's internal parameters are public, so anyone can download and run it. It is described as the open-weight counterpart to Qwen3.8-Max, and "dense" means it uses all of its parameters for every response.

Three specifications stand out:

• Native image and video input. The model reads photos, charts and video as well as text.
• 262,144-token (262K) context. Context is the length of text the model can read at once, so you can hand it a long report or paper whole and ask questions about it.
• A hybrid 64-layer design. It combines 48 layers of linear attention (Gated DeltaNet, a lighter-weight mechanism) with 16 layers of standard attention. This keeps the working memory for long text (the KV cache) small.

Published benchmarks put it close to frontier level in coding and agent use, meaning autonomous multi-step tasks. One point matters for local use: there is no official GGUF. GGUF is a format that compresses (quantizes) weights for local execution, so every lightweight version of this model was converted by someone in the community.

What is abliteration, and how is it measured?#

Abliteration is a weight-editing method that removes a model's refusal behavior without retraining. Researchers found that refusal corresponds to a fixed direction in the model's internal representation. Editing the weights so that activity no longer flows into that direction removes the habit of stopping, while the knowledge stays in place. For the mechanism in detail, see our explainer on what abliteration is.

A standard AI learns during training to stop before answering when dangerous words appear. In 2024, this behavior was shown to correspond to a specific direction on the model's internal "map of meaning." Because knowledge lives elsewhere on that map, only the habit of stopping disappears, and the model's general ability is damaged less than with retraining.

The usual measure of how far an edit has moved the model is KL divergence. It quantifies how much the edited model's output differs from the original's, and a smaller value means behavior closer to the original. A value around 0.01 is a rough guide for "very close to the original."

How do the four abliterated Qwen3.8-27B editions differ?#

There is no single abliterated Qwen3.8-27B. Each creator uses different tools, edits a different range of the model, and publishes different results, so the four editions are effectively different models. huihui-ai narrows its edit range to protect performance, orcarouter leads on speed of release and quantization choices, and RVN pushes KL divergence down.

What does the huihui-ai edition offer?#

huihui-ai is a model-editing workshop. Its edition was published on September 8, 2026, as huihui-ai/Huihui-Qwen3.8-27B-abliterated and inherits the Apache 2.0 license. It was made with a tool called remove-refusals-with-transformers, and the model card calls the result "crude, proof-of-concept."

The latest version edits only layers 18–51. The earlier layers are left untouched to preserve the original performance, and the MTP component (used for speculative decoding, which speeds up responses) and the vision part are also unmodified. The older version can be downloaded at revision d42ca89. GGUF versions exist too, and the variants named UD and UD-DW modify only layers 18–51 and 23–51 respectively. On Ollama, it is available as huihui_ai/Qwen3.8-abliterated.

What makes the orcarouter edition distinct?#

orcarouter was the fastest to publish. It released Qwen3.8-27B-Uncensored-FP8 on August 15, the day after the base model, and the GGUF version on August 16. Its strength is the range of quantizations: F16 plus 12 levels from Q2_K to Q8_0, including imatrix variants such as IQ3_M and IQ4_XS.

The representative Q4_K_M file is 16.8GB, and a 24GB card is recommended. The vision projector is a separate 931MB file, and the KV cache at 8K context is about 0.5GB. The published figures say refusals to harmful prompts fell from 64–99% to 0–6%, over-refusal of harmless questions fell from 5.6% to 0.4%, and performance scores changed by within ±1%. The edition is marked research-only and is not meant for deployment to end users.

What is the trohrbaugh heretic-ara edition?#

This edition was made with the ARA method in the abliteration tool Heretic. It is reported to have a KL divergence of 0.0535 and to refuse 3 of 100 test prompts. It is less a model to use on its own than the base for the RVN edition below.

How does the slevinw RVN edition get its numbers so low?#

The RVN edition, published as Qwen3.8-27B-Heretic-Abliterated-Uncensored-GGUF, applies two more passes of editing on top of the trohrbaugh edition. The result is reported as a KL divergence of 0.0085 and 0–1 refusals per 100 prompts. The one remaining refusal is on a chemical-weapons topic and is said to be left in deliberately.

File sizes are 15.41GiB for Q4_K_M, 11.24GiB for Q3_K_S (aimed at 16GB setups), and 26.63GiB for Q8_0. The multilingual build includes the official chat template and supports speculative decoding with MTP.

Taken together, huihui-ai limits layers to keep performance, orcarouter offers early release and a wide choice of quantizations, and RVN aims for the smallest KL divergence. Because each edition differs in tools, edit range and published figures, the same base name does not mean the same behavior.

What do you need to run abliterated Qwen3.8-27B locally?#

You need about 16–24GB of VRAM, the memory on your GPU. The 27B class is heavy as an entry size. You can pull the huihui-ai edition through Ollama with huihui_ai/Qwen3.8-abliterated, and GGUF versions are also published.

To see which model sizes fit your PC, read our local LLM guide. For how to choose among quantization levels, see our quantization explainer.

What should you know before using an abliterated model?#

What is Deepent Research?#

Deepent Research is a browser-accessible LLM chat service built to refuse as little as possible. It keeps going with research that a standard AI would cut off mid-answer.

• It refuses nothing. We run uncensored and abliterated models with guardrails reduced as far as possible, so prompts a standard AI answers with "I can't help with that" still get a response (Qwen3.5 / Qwen3.8-27B-abliterated, depending on your plan).
• Your input is never used for training. What you look up is used only to provide the service, never to train AI or for any other purpose. Personal data is protected, and our privacy policy is public.
• Try it free. No GPU and no install, just open a browser. In the logged-out trial your conversation isn't saved, so you can start casually.

→ Try it free, no sign-up: https://deepentresearch.com

No question can stop research.

Frequently asked questions#

Is Qwen3.8-27B abliterated an official release from Alibaba?#

⌄

No. Alibaba published the base Qwen3.8-27B as an open-weight model under Apache 2.0, and every abliterated edition is a community modification. Each creator uses different tools and edits different parts of the model, so editions sharing the Qwen3.8-27B name can differ a lot. Alibaba also publishes no official GGUF files, so every quantized edition is a community conversion.

How much do different abliterated Qwen editions vary in practice?#

⌄

They differ in tools, edit range and published numbers, so in practice they behave as separate models. For example, huihui-ai edits only layers 18–51 to protect performance, while the RVN edition reports a KL divergence of 0.0085 and 0–1 refusals per 100 prompts. Check who distributes an edition and which revision you are downloading before you choose.

Can you run Qwen3.8-27B abliterated on your own PC?#

⌄

Yes, if your GPU has enough video memory. The rough guide is 16–24GB of VRAM, and because this is a 27B-class model it is heavy as an entry-level size. You can pull the huihui-ai edition through Ollama as huihui_ai/Qwen3.8-abliterated, and GGUF versions are also published for local use.

Does an abliterated model guarantee accurate answers?#

⌄

No. An abliterated model has lost the mechanism that checks before it answers, so it will not stop on risky questions, but refusing less does not make an answer correct. Judging what the model says is the user's responsibility, so verify important claims against reliable sources before you rely on them.

Disclaimer#

This article was drafted with AI assistance and then reviewed, fact-checked, and edited by the Deepent Research team before publication.
It is provided for educational and research purposes and is not intended to encourage unlawful activity or crime.
Where figures are cited, they come from third-party public reports and vendor-published values and may vary with measurement conditions.