Skip to content

About

Source-tracked catalog of abliterated, low-refusal, and community-reported open-weight and hosted language models

Topics

Resources

Contributing

Stars

19 stars

Watchers

0 watching

Forks

Latest commit

 

History

6 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 

Repository files navigation

Awesome Uncensored LLMs

A source-tracked guide to open-weight, low-refusal, abliterated, and community-reported hosted language models. Entries distinguish model weights from the service or interface used to run them.

“Uncensored” is not a standardized technical property or a guarantee. Abliteration, fine-tuning, prompting, provider routing, and interface policy can all change refusal behavior. This list documents evidence and provenance instead of treating the label as a promise.

README reviewed: 2026-10-06. Individual model availability and provenance still require checking against the linked source.

Related Projects

Contents

How to read this list

This is a shortlist for evaluation, not a benchmark ranking. The first group includes recent hosted model names and variants that may not have a public downloadable checkpoint. A model family name or API alias is not proof that a particular provider serves the same weights. Availability, pricing, context limits, and behavior change; verify the exact endpoint before use. The second group contains open-weight checkpoint candidates with public model cards.

Status Meaning
Verified low-refusal A maintainer reproduced the result on the exact version and recorded the setup.
Community-reported A specific source reports low-refusal behavior; independent reproduction is pending.
Hosted variant The name refers to a hosted fine-tune or deployment; weight provenance may not be public.
Open-weight candidate Weights are available, but that alone does not establish low-refusal behavior.
To verify The entry is a candidate only; claims, access, or license need further evidence.

Current hosted and community-reported candidates

The following candidates were selected for capability, recency, family coverage, or value. They are not ranked against one another. Thinking modes and alternate serving routes are noted as variants rather than counted as separate model families.

Model / variant Family or base Access / status Evaluation focus Primary family source Muapi landing page
Abliterated Model Large V2 GLM-5.3-derived Hosted variant; verify provenance and availability Large reasoning, tools, long context GLM model family Abliterated Model API
GLM 5.3 Flash Uncensored GLM-5.3 Flash Hosted variant; community-reported Fast reasoning, coding, tools, vision where supported Z.AI models GLM Abliterated API
Qwen 3.8 27B Uncensored Qwen 3.8 27B Hosted fine-tune; verify checkpoint General chat, coding, reasoning, multimodal Qwen Qwen Abliterated API
GLM 5.3 Uncensored GLM-5.3 Hosted variant; experimental Full-size reasoning, coding, tool use Z.AI models GLM Abliterated API
Qwen 3.8 27B Obliterated Qwen 3.8 27B Hosted variant; verify method Compare refusal-direction ablation with fine-tunes Qwen Qwen Abliterated API
MiMo V2.6 Flash Uncensored MiMo V2.6 Flash Hosted fine-tune; availability varies Fast chat, reasoning, tool use Xiaomi MiMo MiMo Abliterated API
MiMo V2.6 Flash Abliterated MiMo V2.6 Flash Hosted variant; availability varies Refusal-direction ablation, reasoning, tools Xiaomi MiMo MiMo Abliterated API
Qwen 3.8 27B Uncensored (TEE route) Qwen 3.8 27B Uncensored Hosted route; route-specific privacy claims Compare serving and privacy properties Qwen Qwen Abliterated API
Abliterated Model GLM-derived multimodal Hosted variant; verify provenance Multimodal input, structured output Z.AI models Abliterated Model API
Abliterated Model Large GLM-5.2-derived Hosted variant; verify provenance Large reasoning, tools, long context Z.AI models Abliterated Model API
Gemma 4 26B A4B Uncensored Gemma 4 26B A4B Hosted fine-tune; verify checkpoint MoE reasoning, coding, multimodal Gemma Gemma Abliterated API
Gemma 4 26B A4B Uncensored (TEE route) Gemma 4 26B A4B Uncensored Hosted route; route-specific privacy claims Compare route and modality support Gemma Gemma Abliterated API
Gemma 4 31B Gembrain Uncensored Heretic Gemma 4 31B Community fine-tune; verify checkpoint General chat and reasoning Gemma Gemma Abliterated API
Qwen 3.8 27B Queen Qwen 3.8 27B Hosted creative fine-tune Roleplay and image-aware dialogue Qwen Qwen Character Chat API
Qwen 3.8 27B Fable Qwen 3.8 27B Hosted creative fine-tune Storytelling and character work Qwen Qwen Character Chat API
Llama 3.3 70B Instruct Abliterated Llama 3.3 70B Open-weight checkpoint General chat; older but established family Model card Llama Abliterated API
Qwen2.5 32B Instruct Abliterated Qwen2.5 32B Open-weight checkpoint General chat and coding; legacy comparison Model card Qwen Abliterated API
DeepSeek R1 Distill Llama 70B Abliterated DeepSeek R1 Distill Llama 70B Open-weight checkpoint Reasoning-focused comparison Model card DeepSeek R1 Abliterated API
DeepSeek R1 Distill Qwen 32B Abliterated DeepSeek R1 Distill Qwen 32B Open-weight checkpoint Smaller reasoning-focused option Model card DeepSeek R1 Abliterated API
NeuralDaredevil 8B Abliterated Llama-family 8B Open-weight checkpoint Lightweight local evaluation Model card Llama Abliterated API

Additional hosted variants available through Muapi (2026-09-29)

Muapi currently exposes the following deployed endpoint aliases. The names identify Muapi routes; they do not independently establish the exact upstream derivative, license, or refusal behavior. Those details remain To verify where direct sources or reproducible evaluations are unavailable.

Model / variant Family or base Muapi endpoint slug Status / focus Primary family source Muapi landing page
GLM 5.3 Flash Abliterated GLM 5.3 Flash glm-5-3-flash-abliterated Muapi hosted endpoint; derivative provenance and behavior to verify Z.AI models GLM Abliterated API
GLM 5.3 Abliterated GLM 5.3 glm-5-3-abliterated Muapi hosted endpoint; derivative provenance and behavior to verify Z.AI models GLM Abliterated API
Gemma 4 26B A4B Abliterated Gemma 4 26B A4B gemma-4-26b-a4b-abliterated Muapi hosted endpoint; derivative provenance and behavior to verify Gemma Gemma Abliterated API
Gemma 4 31B Gembrain Abliterated Gemma 4 31B gemma-4-31b-gembrain-abliterated Muapi hosted endpoint; checkpoint provenance and behavior to verify Gemma Gemma Abliterated API
Qwen 3.8 27B Abliterated Qwen 3.8 27B qwen-3-8-27b-abliterated Muapi hosted endpoint; ablation method and behavior to verify Qwen Qwen Abliterated API
GLM 4.6 Derestricted v5 GLM 4.6 glm-4-6-derestricted-v5 Muapi hosted endpoint; derivative provenance and behavior to verify Z.AI models GLM Abliterated API
Qwen3.5 27B Queen Derestricted Qwen3.5 27B qwen-3-5-27b-queen-derestricted Muapi hosted endpoint; derivative provenance and behavior to verify Qwen Qwen Abliterated API
Qwen3.5 27B Blossom Derestricted Qwen3.5 27B qwen-3-5-27b-blossom-derestricted Muapi hosted endpoint; derivative provenance and behavior to verify Qwen Qwen Abliterated API
Qwen3.5 27B Opus-Distilled Derestricted Qwen3.5 27B qwen-3-5-27b-opus-distilled-derestricted Muapi hosted endpoint; distillation provenance and behavior to verify Qwen Qwen Abliterated API
Gemma 4 31B SDFT Abliterated Gemma 4 31B gemma-4-31b-sdft-abliterated Muapi hosted endpoint; checkpoint provenance and behavior to verify Gemma Gemma Abliterated API
Venice Abliterated Venice venice-abliterated Muapi hosted endpoint; upstream model and behavior to verify Venice AI Low-Refusal LLM API

These aliases may overlap with existing “uncensored” or creative fine-tune entries in family, but they are kept separate because the endpoint names claim different modifications or serving routes. Do not merge them until equivalence is established. The Muapi endpoint slugs are included to identify the deployed routes; consult the Abliterated LLM API page for current access details.

Important provenance note

Several recent hosted names in the table do not have a linked public derivative model card in this catalog. Their base-family links are not proof of the derivative's provenance, license, or behavior. Keep them marked as hosted/community-reported until a direct source and reproducible evaluation are available. Do not infer a model's local weights from an API name.

Open-weight checkpoints

Start with the exact model cards linked above. Review each card's license, base-model terms, quantization, and inference template before downloading or redistributing. Abliteration changes weights and behavior; it does not grant rights beyond the original model license.

Use cases and choosing an access path

This catalog includes different kinds of access and model families. Choose by what you need to do, then verify the exact model and version; a low-refusal report is not a quality score.

Use case Access path to consider Checks before use
Try a hosted model without managing inference A hosted endpoint in the Muapi abliterated LLM guide or another listed provider. Exact route/model ID, current availability, price, accepted inputs, output limits, provider terms, and whether the route is a hosted alias.
Run locally or offline An open-weight checkpoint from the table above. The exact checkpoint card, base and derivative licenses, hardware needs, quantization, inference template, and model behavior on your prompts.
Compare a modified model with its base A checkpoint and its upstream base, or a hosted alias with a supported baseline. Same prompt, system message, sampling settings, inference path, and version where possible; record any differences you cannot control.
Evaluate coding or repository work Coding-oriented candidates plus a coding-agent integration when tools are required. Separate model-only code answers from agent tool use. See Awesome Uncensored Coding Models.
Use a general-purpose agent A model with verified tool calling connected to a bounded agent stack. Model capability, endpoint compatibility, agent permissions, confirmation gates, and spend limits. See Awesome Uncensored AI Agents.
Explore image or video generation The relevant companion image or video catalog, not a text model entry. Exact modality, checkpoint/endpoint, content rules, model license, and whether the claim applies to local weights or a hosted pipeline.
Study refusal behavior Exact base/modified pairs or provider routes with reproducible setup details. Define benign and safety-control prompts in advance; keep refusal rate separate from correctness, calibration, and task quality.

A small evaluation plan

  1. Choose a few tasks representative of your intended use, including ordinary tasks and edge cases.
  2. Record the exact model/checkpoint or route, source, access type, version/date, runtime or provider, prompt, and settings.
  3. Use the same setup for a baseline where possible; report unavoidable differences.
  4. Run multiple samples when outputs vary, preserve raw outputs, and include errors or unavailable models.
  5. Score factuality, task completion, instruction following, refusal behavior, latency, and cost as separate measures.
  6. Review model and provider terms before sending private data or relying on a hosted route.

This list is a source-tracked directory, not a certification of models and not a leaderboard. An entry should be treated as a candidate until its exact version and claims have been checked.

Selection notes

  • Run the same refusal and capability prompts against the base and modified model, with the model version, sampler, system prompt, and inference stack recorded.
  • Test harmless edge cases as well as ordinary tasks; a low refusal rate can come with degraded calibration or instruction following.
  • Separate refusal behavior from factuality, quality, latency, and tool reliability.
  • For hosted models, record the exact provider, route, model snapshot, region, date, and terms.
  • Mark a model unavailable when the endpoint cannot be selected or returns an availability error; do not silently replace it with a similarly named model.

Contributing

See CONTRIBUTING.md. Add primary sources, exact version identifiers, access method, licensing details, and reproducible evaluation notes. Do not make unsupported “uncensored” claims.

Responsible use

Use these resources lawfully and responsibly. Do not create or facilitate sexual content involving minors, non-consensual intimate imagery, targeted harassment, fraud, impersonation, or other illegal or abusive material. Follow model licenses, service terms, and applicable law.

License

The original catalog and documentation are licensed under CC BY 4.0. Individual models, code, datasets, checkpoints, APIs, and linked resources retain their own licenses and terms.

About

Source-tracked catalog of abliterated, low-refusal, and community-reported open-weight and hosted language models

Topics

Resources

Contributing

Stars

19 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors