Ollama Uncensored and Abliterated Models: Picks by VRAM (November 2026)

Short answer: on Ollama the abliterated builds live in the huihui_ai namespace, and the right one is set by your VRAM. 8GB: huihui_ai/qwen3-abliterated:8b (5.0GB), or dolphin3:8b (4.9GB) if you would rather have a filtered fine-tune than weight surgery. 12GB: huihui_ai/gemma-4-abliterated:12b (7.6GB, 128K context, text and image input) — the most-pulled abliterated line on Ollama at 1.3M pulls. 16GB: huihui_ai/qwen3-abliterated:14b (9.0GB). 24GB: huihui_ai/qwen3.5-abliterated:27b (17GB, 256K) or the two-month-old huihui_ai/Qwen3.8-abliterated:27b (18GB). 48GB: huihui_ai/llama3.3-abliterated:70b (43GB). Every tag and size on this page was read from ollama.com on 2 November 2026.

Two things before the list, because most pages on this topic skip both. First, "uncensored" covers two very different techniques — a dataset-filtered fine-tune and a weight-surgery trick called abliteration — and they have different quality costs. Second, those costs are real: abliterated models are measurably worse at reasoning than their stock versions, and we will say so plainly rather than pretend otherwise. If you just want our general Ollama rankings with filters intact, that is the best Ollama models guide; this page is for when the stock model refuses work you have every right to do — dark fiction, security research, medical or legal questions phrased bluntly, villain dialogue that a corporate safety layer keeps rewriting into an apology.

---

What "Uncensored" Actually Means

Two techniques produce "uncensored" models: filtering alignment out of the training data (Dolphin's approach), or abliteration — surgically deleting the refusal direction from a finished model's weights. They are not interchangeable, and knowing which one you are pulling predicts how the model will behave.

Dataset-filtered fine-tunes. Eric Hartford's Dolphin series takes a strong base model and fine-tunes it on instruction data that has been filtered to remove alignment and moralizing responses. The model never learns the refusal reflex in the first place. Because this is a normal fine-tune, quality holds up well — Dolphin models behave like polished assistants that simply do not lecture you.

Abliteration. A finished, aligned model gets its weights edited directly. The insight behind it: refusal behavior in chat models is largely mediated by a single direction in the model's activation space. Find that direction, subtract it out of the weights, and the "I can't help with that" reflex disappears while the knowledge stays. It is fast and cheap — which is why abliterated builds of new models appear on Ollama within weeks of release — but it is surgery on a live patient. The most prolific publisher, huihui_ai, describes its own method on the qwen3-abliterated page as "a crude, proof-of-concept implementation to remove refusals from an LLM model", built on the open-source remove-refusals-with-transformers project. Respect that honesty; it matches how the models behave.

What neither technique does: add knowledge the base model never had, or make output legal that was not legal before. An uncensored model is the same model with the "no" removed — nothing more.

---

The Dolphin Family: The Uncensored Default

If you want one uncensored model and no research project, pull dolphin3:8b. The Dolphin line has been the reference uncensored fine-tune since the Llama 2 days, and the 3.0 generation is still the one to pull — nothing newer has appeared in the official library since it shipped a year ago.

Speed matches the base weights, because the download is the same size as the stock model: a 4.9GB file is a 4.9GB file whether or not it was fine-tuned on filtered data. The ceiling is your card's memory bandwidth divided by that file — about 90 tok/s for dolphin3:8b on an RTX 5060 Ti 16GB (448 GB/s), about 200 tok/s on an RTX 4090 (1,008 GB/s) — and real runs come in under it. The only measured numbers on this site are the M3 Pro table in the DeepSeek R1 guide, where the 7B–8B Q4 files land at 80–90% of their ceiling.

One thing worth quoting from Dolphin's own model card, because it frames the whole category correctly: the model is steerable and gives control to the system owner — you set the system prompt, you decide the alignment — and the authors explicitly advise implementing your own alignment layer before exposing it as a service. In other words: built for you running it on your machine, not for you wiring it into a public chatbot unfiltered.

---

Abliterated Builds of Current Models

When you want this year's base models without refusals, the huihui_ai namespace on Ollama is where they live — abliterated builds of Gemma 4, Qwen 3.5, Qwen 3.6, Qwen 3.8, gpt-oss, Qwen 3, Llama 3.3 and DeepSeek-R1, published as ordinary pullable tags. The namespace page lists over a hundred lines; these are the ones that matter for a consumer card, with the sizes read from each tag page on 2 November 2026.

At the same quantization, footprints match the stock models — abliteration edits weights, it does not shrink them — so the sizing rules from the Ollama RAM/VRAM table carry over unchanged, and so does the speed arithmetic: the ceiling is memory bandwidth ÷ download, so the 9.0GB qwen3-abliterated:14b tops out near 50 tok/s on an RTX 5060 Ti 16GB (448 GB/s) and near 112 tok/s on an RTX 4090 (1,008 GB/s), with real runs below that. The catch is that huihui_ai does not publish every line at Q4: gemma3-abliterated only exists at q8_0 and fp16, and Qwen3.8-abliterated ships a dozen quants from q2_K (12GB) to bf16 (56GB) under one 27b label. Open the tags page before you pull.

Two notes on what you are getting. The newer huihui_ai pages (gemma-4, qwen3.5, Qwen3.8) no longer carry the "crude, proof-of-concept" line — they link the remove-refusals-with-transformers project and add a usage-warning block that recommends the models for "research, testing, or controlled environments" rather than "production or public-facing commercial use". And other namespaces publish "uncensored" and "heretic" builds of the same bases (eriktonon, mgraffam, abliterant and others appear in the Ollama search), mostly with a few thousand pulls each; this page sticks to huihui_ai and the official library because those are the tag pages we checked.

The trade: these are the newest, smartest bases you can get uncensored, and also the builds where the quality cost of the surgery shows most. More on that below — read the trade-off section before you make one of these your daily driver.

---

Hermes 3: The Steerable Middle Ground

Hermes 3 (ollama pull hermes3:8b, 4.7GB) is not marketed as "uncensored" — it is neutrally aligned, which in practice means the system prompt decides everything. Nous Research fine-tuned Llama 3.1 to follow instructions exactly and neutrally, without refusal behavior baked into the weights; the behavioral boundaries are whatever you write in the system prompt.

For a lot of people this is actually the right answer. Nothing was filtered out and nothing was ablated, so Hermes 3 keeps essentially all of its base-model quality — it is a genuinely strong generalist with good roleplay, reasoning and long-context coherence, sitting in the official Ollama library at 3b (2.0GB) and 8b (4.7GB), with 70b (40GB) and 405b (229GB) for bigger hardware, all at 128K context. Write a system prompt that says who the model is and what it will and will not do, and it follows that — yours, not Meta's.

The honest caveat: neutral alignment is not zero alignment. With an empty system prompt Hermes 3 is more conservative than Dolphin or an abliterated build. It rewards people who actually use system prompts; if you never touch them, Dolphin gives you the uncensored default with less setup.

---

The Llama-2-Era Classics — and Why to Skip Them

llama2-uncensored, wizard-vicuna-uncensored and wizardlm-uncensored still sit in the Ollama library — all three last updated three years ago — and in 2026 we would not install any of them. These early builds invented the category — Eric Hartford's Wizard pair plus George Sung and Jarrad Hope's llama2-uncensored — and wizard-vicuna-uncensored (7b/13b/30b) was the uncensored model for a generation of local AI users. But they are built on Llama 2-era bases with short contexts and reasoning that a modern 8B beats without trying.

They show up in every "uncensored Ollama" listicle because those listicles copy each other, and the pull counts (llama2-uncensored is at 2.8M) are mostly that history. The only reason to pull one today is curiosity about where the scene came from. For actual use, dolphin3:8b is smaller than the old 13Bs and better at everything.

---

VRAM: What Fits Your Card

The 8B-class picks download at 4.9–5.0GB, the 12B Gemma 4 at 7.6GB, the 14B at 9.0GB, the 27Bs at 17–18GB and the 70B at 43GB — at the same quant, uncensored builds cost the same memory as their stock versions (just check the tag, since not every abliterated line is published at Q4). Match the pick to your card:

The usual fitting rules apply — Q4_K_M is the sweet spot, and it is the KV-cache growing with context that triggers out-of-memory errors, not the weights. Our tier guides cover the details for your card: 8GB, 12GB, 16GB and 24GB, and the Ollama RAM/VRAM table has the download size for every common tag.

---

The Quality Trade-Off: What You Give Up

Abliteration makes models measurably worse — sloppier instruction-following and degraded multi-step reasoning — and anyone selling you an abliterated build as a free lunch is wrong. The refusal direction does not live in clean isolation; subtracting it out of the weights nicks the capabilities stored near it. How much varies by model and by how aggressively the ablation was run, which is why huihui_ai keeps re-releasing improved versions of the same abliterations (the qwen3 and gemma-4 lines both carry -v2 tags).

The practical hierarchy, from least to most quality lost:

1. Hermes 3 — nothing removed, nothing ablated. Essentially full base-model quality; you pay in system-prompt effort instead.
2. Dolphin — a normal fine-tune on filtered data. Holds up well; the cost is that its base (Llama 3.1) now trails the newest releases.
3. Abliterated builds — newest bases, real surgery. Fine for conversation and creative work; noticeably less reliable on hard multi-step reasoning and precise formatting than their stock versions.

Our advice: keep a stock model installed next to your uncensored one and switch per task. Uncensored for the writing, roleplay and blunt questions the stock model refuses; stock for the code review and the reasoning-heavy work. Ollama makes holding both a non-issue — a second 5GB model costs you disk, not VRAM, until you load it.

---

Frontends: Where the Creative-Writing Crowd Runs These

Ollama runs every model on this page, but for long-form fiction and roleplay most of this community runs KoboldCpp or text-generation-webui instead — same GGUF weights, better tooling for the job. KoboldCpp adds character cards, world info and memory management that a chat UI does not have; text-generation-webui exposes the sampler controls that matter when you are steering a model through a 20K-token story. We keep setup guides for both: KoboldCpp and text-generation-webui.

And if you came here for the image side of this question — running FLUX and SDXL without content filters — that is its own page: uncensored local image generation.

---

Responsible Use: Local Is Not Lawless

Running these models is legal, and the legitimate reasons to run them are real — fiction with actual villains, security research, unfiltered answers to medical and legal questions, and simply not having a corporation's content policy sitting between you and your own hardware. But be clear-eyed about what you have removed: a safety layer, not a law. Whatever is illegal to produce or possess in your jurisdiction remains exactly as illegal when a local model generates it, and you are the publisher of anything you do with the output. There is no place on your machine or anyone else's for content that exploits real people or breaks the law — and the models' own authors say the same thing. Dolphin's card tells you to add your own alignment layer before exposing it as a service; huihui_ai's pages say the builds are for research, testing or controlled environments, not public-facing use, and that the user is responsible for complying with local law. "Uncensored on my desk" and "unfiltered on the public internet" are different decisions. Run them privately, keep responsibility for the output, and this whole category stays what it should be: a tool for adults.

---

Sources

• Ollama library — dolphin3, dolphin-llama3, dolphin-mistral, dolphin-mixtral, hermes3, llama2-uncensored, wizard-vicuna-uncensored, wizardlm-uncensored (tags and sizes read 2 November 2026)
• huihui_ai on Ollama — the namespace listing, plus the tag pages for gemma-4-abliterated, qwen3.5-abliterated, Qwen3.6-abliterated, Qwen3.8-abliterated, gpt-oss-abliterated, qwen3-abliterated, gemma3-abliterated, deepseek-r1-abliterated, llama3.3-abliterated and dolphin3-abliterated
• remove-refusals-with-transformers — the abliteration implementation every huihui_ai page links
• Nous Research — Hermes 3 technical report — neutral-alignment design
• Cognitive Computations (Eric Hartford) — Dolphin model cards and dataset-filtering methodology
• Bandwidth ceilings: NVIDIA GeForce compare table (RTX 50-series memory bandwidth) and the NVIDIA Ada architecture whitepaper (RTX 4090, 1,008 GB/s); ceiling = bandwidth ÷ Ollama download size

---

FAQ