Uncensored models · checked Oct 23, 2026
Uncensored and low-refusal models.
Open-weight and hosted models with fewer refusals, scored on the UGI benchmark, with where to run each one, the VRAM you need, the license and every source.
#1 UGI here: GLM-4.6 Derestricted v3 (69.82)12 models · 1 toolUGI board 2026-10-16
UGI benchmark
UGI Leaderboard (Uncensored General Intelligence): UGI (uncensored knowledge+willingness), W/10 (willingness 0 to 10), NatInt (general intelligence). 1,329 rows on the board; latest test date 2026-10-16.
Swipe the chart sideways for all 8 models →
UGI score (uncensored knowledge + willingness) · 8 models · top 5 coloured · axis does not start at zeroSource: UGI Leaderboard ↗ · board dated 2026-10-16, checked 2026-10-23
Not on UGI yet: Venice Uncensored 1.2 / Venice Role Play Uncensored (hosted), Huihui Qwen3.8 27B abliterated, Qwen3.8 27B Uncensored (GGUF), Wicked-Tutu-31B, Chroma1-HD, Venice image models (hosted). Many 2026 community models (Qwen3.8-27B uncensored/abliterated, Wicked-Tutu) are not on UGI yet.
At a glance
Text (LLM)Dolphin Mistral 24B Venice EditionUGI 40.79Text (LLM)Venice Uncensored 1.2 / Venice Role Play Uncensored (hosted)Apache-2.0 for the open Role Play weights; hosted API under Venice termsText (LLM)Hermes 4 / 4.3 (Nous Research)UGI 54.07Text (LLM)Huihui Qwen3.8 27B abliteratedQ4 file 16.6 GBText (LLM)Qwen3.8 27B Uncensored (GGUF)Q4 file 16.8 GBText (LLM)Gemma 4 31B it HereticUGI 65.69ToolHeretic (tool)AGPL-3.0Text (LLM)Llama 3.1 8B / Gemma 3 27B abliterated (mlabonne)UGI 41.05Text (LLM)Wicked-Tutu-31BQ4 file 18.7 GBText (LLM)TheDrummer fine-tunes (Behemoth, Artemis, Orion)UGI 57.19Text (LLM)GLM-4.6 Derestricted v3UGI 69.82ImageChroma1-HDApache-2.0ImageVenice image models (hosted)hosted API
Where to run each model
Licence, minimum VRAM, UGI score and sources with dates. VRAM figures are estimates from Q4_K_M file sizes.
| Model | Type | License | Where to run it | Min VRAM | UGI | Sources |
|---|---|---|---|---|---|---|
| Dolphin Mistral 24B Venice EditionDolphin (dphn, formerly Cognitive Computations) with Venice.ai · 24B · released 2025-06-12Base: mistralai/Mistral-Small-24B-Instruct-2501 | Text (LLM)fine-tune (uncensored instruct) | Apache-2.0 | Hugging Face ↗GGUF ↗LM Studio GGUF (search the repo name)Hosted ↗ Venice API id venice-uncensored-1-2 is a newer hosted Venice Uncensored version | ~14.3 GB file at Q4_K_M; plan on a GPU with at least 17 GB VRAM (or unified memory) for short contexts | 40.79 W/10 7.8 · 2025-10-04 | checked 2026-10-23 |
| Venice Uncensored 1.2 / Venice Role Play Uncensored (hosted)Venice.ai · 24B (Role Play, per HF) · released 2026-05-29 (Role Play weights on HF)Base: Venice Role Play Uncensored is a fine-tune of Dolphin-Mistral-24B-Venice-Edition (open weights on HF); Venice Uncensored 1.2 base not stated | Text (LLM)hosted API | Apache-2.0 for the open Role Play weights; hosted API under Venice terms | hosted (Role Play weights: same size class as the 24B Dolphin model) | Not on UGI yet | docs.venice.ai ↗ 2026-10-23huggingface.co ↗ 2026-10-23 checked 2026-10-23 | |
| Hermes 4 / 4.3 (Nous Research)Nous Research · 14B / 36B / 70B / 405B · released 2025-08 (Hermes 4), 2025-11-17 (Hermes 4.3 36B)Base: Hermes-4.3-36B: ByteDance-Seed/Seed-OSS-36B-Base; Hermes-4-14B: Qwen/Qwen3-14B; Hermes-4-70B/405B: Llama 3.1 | Text (LLM)fine-tune (low-refusal, steerable) | Apache-2.0 (4.3-36B, 4-14B per HF); 70B/405B follow the Llama 3.1 license | Hugging Face ↗Ollama ↗ Hermes 3 only, 1.8M downloads, updated 1 year agoHosted Venice API lists hermes-3-llama-3.1-405b | 36B at 4-bit is roughly 20+ GB; 14B fits ~10 GB | Hermes-4.3-36B: 38.06 W/10 6.2Hermes-4-405B: 54.07 W/10 4.8Hermes-4-70B: 34.01 W/10 3.2 | huggingface.co ↗ 2026-10-23huggingface.co ↗ 2026-10-23 checked 2026-10-23 |
| Huihui Qwen3.8 27B abliteratedhuihui-ai (community) · 27.3B · released 2026-08-16Base: Qwen/Qwen3.8-27B | Text (LLM)abliteration (refusal direction removed) | Apache-2.0 | ~16.6 GB file at Q4_K_M; plan on a GPU with at least 20 GB VRAM (or unified memory) for short contexts | Not on UGI yet | checked 2026-10-23 | |
| Qwen3.8 27B Uncensored (GGUF)JonathanColetti (community); variants by HauhauCS, 0bserverx (Heretic) and others · 27.3B · released 2026-08-14Base: Qwen/Qwen3.8-27B | Text (LLM)uncensored fine-tune / abliteration (community) | Apache-2.0 | ~16.8 GB file at Q4_K_M; plan on a GPU with at least 20 GB VRAM (or unified memory) for short contexts | Not on UGI yet | checked 2026-10-23 | |
| Gemma 4 31B it Hereticcoder3101 (community) · 31B · released 2026-04-02Base: google/gemma-4-31B-it | Text (LLM)Heretic (automatic abliteration) | Apache-2.0 (as tagged on HF) | 31B at 4-bit is roughly 19 GB file; 24 GB GPU | 65.69 W/10 10 · 2026-04-12 | huggingface.co ↗ 2026-10-23huggingface.co ↗ 2026-10-23 checked 2026-10-23 | |
| Heretic (tool)p-e-w (GitHub) · released 2025-09-21Base: works on most transformer LLMs | Toolautomatic censorship removal (directional ablation with parameter search) | AGPL-3.0 | Not stated | Not on UGI yet | api.github.com ↗ 2026-10-23huggingface.co ↗ 2026-10-23 checked 2026-10-23 | |
| Llama 3.1 8B / Gemma 3 27B abliterated (mlabonne)Maxime Labonne (community) · 8B / 27B · released 2024-07-24 / 2025-03-16Base: meta-llama/Meta-Llama-3.1-8B-Instruct; google/gemma-3-27b-it | Text (LLM)abliteration | Llama 3.1 license / Gemma terms | 8B: 4.9 GB Q4_K_M (6 GB GPU); 27B: 16.5 GB Q4_K_M (20 GB+) | gemma-3-27b-it-abliterated: 41.05 W/10 8.2 | huggingface.co ↗ 2026-10-23huggingface.co ↗ 2026-10-23 checked 2026-10-23 | |
| Wicked-Tutu-31BReadyArt (community) · 31B · released 2026-10-16 (v0.1); Final GGUF 2026-10-20Base: google/gemma-4-31B-it | Text (LLM)fine-tune (rank-128 LoRA per the promo post) for creative writing | Apache-2.0 (as tagged on HF) | Hugging Face ↗LM Studio GGUF | ~18.7 GB file at Q4_K_M; plan on a GPU with at least 22 GB VRAM (or unified memory) for short contexts | Not on UGI yet | checked 2026-10-23 |
| TheDrummer fine-tunes (Behemoth, Artemis, Orion)TheDrummer (community) · 26B to 128B · released Artemis 31B v1.2 2026-09-25; Behemoth-X-123B v2.2 2026-10-07Base: Artemis-31B: google/gemma-4-31B-it; others vary | Text (LLM)creative/role-play fine-tunes | not stated on the Artemis-31B-v1.2 card | Not stated | Behemoth-X-123B-v2: 57.19 W/10 8.2 | huggingface.co ↗ 2026-10-23 checked 2026-10-23 | |
| GLM-4.6 Derestricted v3ArliAI (community) · 355B · released 2025-12-02Base: zai-org/GLM-4.6 | Text (LLM)derestricted fine-tune | MIT | server class (355B MoE) | 69.82 W/10 8.8 · 2026-03-02 | huggingface.co ↗ 2026-10-23huggingface.co ↗ 2026-10-23 checked 2026-10-23 | |
| Chroma1-HDlodestones (community) · 8.9B · released 2025-08-08Base: FLUX.1-schnell architecture | Imageopen base model for fine-tuning | Apache-2.0 | 8.9B in bf16 needs ~18 GB+; quantized/offloaded runs fit smaller GPUs | Not on UGI yet | huggingface.co ↗ 2026-10-23huggingface.co ↗ 2026-10-23 checked 2026-10-23 | |
| Venice image models (hosted)Venice.aiBase: various; e.g. qwen-edit-uncensored, chroma, wai-Illustrious, venice-sd35 | Imagehosted | hosted API | Not stated | Not on UGI yet | docs.venice.ai ↗ 2026-10-23 checked 2026-10-23 |
Official X accounts
- @dphnAI ↗ Dolphin lab account, bio 'AI Lab developing unrestricted models'
- @AskVenice ↗
- @NousResearch ↗ verified
Sources and gaps
- UGI Leaderboard (Uncensored General Intelligence) ↗ · board 2026-10-16 · checked 2026-10-23
- No open-weight video model makes an official 'uncensored' claim that I could verify; none listed.
- Several community authors have no verifiable official X account.
- HF download counters are rolling (about 30 days), not lifetime.
- VRAM figures are estimates from Q4_K_M file sizes, not vendor specs.
FAQ
What does "uncensored" mean here?
Models that were fine-tuned or edited (for example by abliteration) to refuse less often than their base model. Some are hosted services that label models uncensored. This page lists facts and links only.
Which open model scores highest on UGI?
GLM-4.6 Derestricted v3 at 69.82 (willingness 8.8/10, tested 2026-03-02) among the models listed here. Board checked 2026-10-23.
How much VRAM do I need?
The figures are estimates from 4-bit (Q4_K_M) file sizes, not vendor specs: plan on a GPU with a few GB more than the file size for short contexts. Server-class models need multi-GPU setups.