[{"data":1,"prerenderedAt":38},["ShallowReactive",2],{"profile-qwen3-8-27b-uncensored":3},{"slug":4,"title":5,"seoTitle":6,"description":7,"summary":8,"published":9,"updated":9,"reviewed":9,"html":10,"toc":11,"sources":25},"qwen3-8-27b-uncensored","Qwen3.8-27B uncensored and abliterated variants","Qwen3.8-27B Uncensored: Abliterated Variants Compared","Qwen3.8-27B uncensored variants: huihui-ai's abliterated weights and HauhauCS's GGUF release, with methods, licenses, hosting, prices and memory needs.","Qwen3.8-27B is the Qwen Team's dense 27B vision-language model under Apache 2.0, with 262,144 tokens of native context. Its best-documented refusal-reduced variant is huihui-ai's abliterated release, which names the layers it changed; HauhauCS publishes a GGUF build without describing its method. Both inherit the base model's license.","2026-11-14","\u003Ch2 id=\"what-the-base-model-is\">What the base model is\u003C\u002Fh2>\n\u003Cp>Qwen3.8-27B is a dense model from the Qwen Team: every one of its roughly 27 billion parameters is used for every token, unlike the mixture-of-experts models elsewhere in this catalogue. Its card describes it as a native vision-language model that reads images and video, with a thinking mode that is on by default and can be turned off per request.\u003C\u002Fp>\n\u003Cp>The architecture is unusual. Of its 64 layers, 48 use Gated DeltaNet, a form of linear attention, and 16 use full attention, in a repeating pattern of three linear layers and one full layer. Linear-attention layers keep a fixed-size state instead of a cache that grows with every token, which is why this model’s KV cache is small for its size. The \u003Ca href=\"\u002Ftools\u002Fvram-estimator\u002F\">VRAM estimator\u003C\u002Fa> counts only the 16 full-attention layers for that reason.\u003C\u002Fp>\n\u003Cp>The card states a native context of 262,144 tokens, “extensible up to 1,000,000”. Hosted providers set their own limits, which in our data range from 32,768 to over a million tokens depending on the provider and the build.\u003C\u002Fp>\n\u003Ch2 id=\"the-variants-at-a-glance\">The variants at a glance\u003C\u002Fh2>\n\u003Cp>\u003Cstrong>huihui-ai’s abliterated weights\u003C\u002Fstrong> are the most transparent option. The card says the release was made with abliteration, a weight edit that removes a “refusal direction” identified in the model’s activations, and that only layers 18 to 51 were edited in the current version. The multi-token-prediction head and the vision encoder were left untouched. An earlier revision kept only the first 15 layers unedited, and the card gives the command to download it. huihui-ai also publishes GGUF files and an Ollama build.\u003C\u002Fp>\n\u003Cp>\u003Cstrong>HauhauCS’s “Aggressive” GGUF release\u003C\u002Fstrong> describes itself as a profile for “direct answers” with minimal preamble, ships the model’s native multi-token-prediction head plus its own speculative-decoding sidecar, and reports “0\u002F465 refusals” on a test it does not publish. It does not say how the change was made. The creator itself suggests a “Balanced” release is the safer default for long-context agentic work when one exists.\u003C\u002Fp>\n\u003Cp>Hugging Face lists many more Qwen3.8-27B variants, including Heretic builds and fine-tunes; several are served by Featherless under their own repo names. We catalogue the two above because their cards are readable and their creators have a record of earlier releases. That is not a judgment that the others are worse.\u003C\u002Fp>\n\u003Ch2 id=\"which-one-for-which-need\">Which one for which need\u003C\u002Fh2>\n\u003Cul>\n\u003Cli>\u003Cstrong>To run locally and inspect what changed:\u003C\u002Fstrong> huihui-ai’s release, in GGUF form. You can compare its outputs with the base model on your own prompts, because the base weights are public under the same license.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>For a hosted API serving known weights:\u003C\u002Fstrong> Featherless serves \u003Ccode>huihui-ai\u002FHuihui-Qwen3.8-27B-abliterated\u003C\u002Fcode> by name. Its published output rate for this model is much higher than its rate for the standard Qwen3.8-27B, and its listed context is 32,768 tokens.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>For a hosted API with long context on one bill:\u003C\u002Fstrong> MolochAI’s \u003Ccode>qwen3.8-27b-uncensored\u003C\u002Fcode> lists a 200,704-token context at its flat rate, but the provider does not say which weights it serves. Venice serves only the standard Qwen 3.8 27B in our data.\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>Prices for every option are in the table below and on the \u003Ca href=\"\u002Fapi\u002Fpricing\u002F#qwen3-8-27b-uncensored\">price table\u003C\u002Fa>.\u003C\u002Fp>\n\u003Ch2 id=\"limits-and-open-questions\">Limits and open questions\u003C\u002Fh2>\n\u003Cp>Neither variant publishes a measurement of capability after the change. Abliteration usually aims to leave knowledge intact, but the research behind it (\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2406.11717\" rel=\"noopener\" class=\"ext\">Arditi et al., 2024\u003C\u002Fa>) and later community work show that results vary by model and by how many layers are edited. If you rely on the vision or thinking features, test them: huihui-ai says the vision part was not modified, but the language layers that interpret images were.\u003C\u002Fp>\n\u003Cp>Apache 2.0 allows commercial use and modification, with notice requirements when you redistribute. That covers the weights; it does not make every use lawful, and responsibility for outputs stays with you. See \u003Ca href=\"\u002Flearn\u002Fllm-model-licenses\u002F\">how the main licenses differ\u003C\u002Fa>.\u003C\u002Fp>\n",[12,16,19,22],{"id":13,"text":14,"level":15},"what-the-base-model-is","What the base model is",2,{"id":17,"text":18,"level":15},"the-variants-at-a-glance","The variants at a glance",{"id":20,"text":21,"level":15},"which-one-for-which-need","Which one for which need",{"id":23,"text":24,"level":15},"limits-and-open-questions","Limits and open questions",[26,30,34],{"title":27,"publisher":28,"url":29},"Qwen\u002FQwen3.8-27B model card","Hugging Face (Qwen Team)","https:\u002F\u002Fhuggingface.co\u002FQwen\u002FQwen3.8-27B",{"title":31,"publisher":32,"url":33},"huihui-ai\u002FHuihui-Qwen3.8-27B-abliterated","Hugging Face (huihui-ai)","https:\u002F\u002Fhuggingface.co\u002Fhuihui-ai\u002FHuihui-Qwen3.8-27B-abliterated",{"title":35,"publisher":36,"url":37},"Refusal in Language Models Is Mediated by a Single Direction","arXiv (Arditi et al., 2024)","https:\u002F\u002Farxiv.org\u002Fabs\u002F2406.11717",1794622336923]