---
title: "Llama 3.2 MoE Uncensored model card & variants"
description: "Compare Llama 3.2 MoE uncensored model variants, capabilities, Hugging Face sources, deployment paths, hosted alternatives, and FAQs."
canonical: https://vynaris.com/uncensored-models/llama-3-2-moe-uncensored
updated: 2026-09-20
---

[Uncensored model directory](https://vynaris.com/uncensored-models) / Llama 3.2 MoE Uncensored

Community model guide Llama 3.2 community MoE

# Llama 3.2 MoE Uncensored

Community mixture-of-experts builds combining Llama-derived experts for creative work, roleplay, and reduced-refusal evaluation.

[View community variants](https://huggingface.co/DavidAU/Llama-3.2-8X3B-MOE-Dark-Champion-Instruct-uncensored-abliterated-18.4B-GGUF) [Compare models](https://vynaris.com/uncensored-models)

- **Scale**: 18B–21B class
- **Architecture**: Community mixture-of-experts builds
- **Modality**: Text
- **Context**: Variant and runtime dependent

Model card

## What is Llama 3.2 MoE Uncensored?

Llama 3.2 MoE uncensored is not one official Meta release. It describes community-built mixtures that combine multiple Llama-derived experts and apply reduced-refusal modifications.

These models are popular with enthusiasts who want more diverse behavior than a single dense checkpoint. Architecture, active experts, prompt format, memory requirements, and quality can differ significantly between releases.

**Directory listing—not a hosted claim**

Vynaris does not currently claim to host this exact model family. Use the linked repositories or compare the clearly labelled hosted alternative.

Capabilities

## Where this family fits

- Creative and conversational generation
- Role-based adversarial testing in authorized environments
- Long-form ideation and scenario generation
- Experimentation with community mixture-of-experts routing

Limitations

## What to verify first

- The family name covers multiple independent community constructions
- Performance and memory estimates cannot be transferred safely between variants
- Compatibility depends on the repository's prescribed runtime and prompt template

Community ecosystem

## Llama 3.2 MoE Uncensored variants on Hugging Face

These are independent community releases. Vynaris is not affiliated with their publishers and does not imply that similarly named variants are identical. Check each repository card and license before use.

DavidAU

### Dark Champion 8x3B

An 18.4B-class community MoE build focused on instruction following and reduced refusals.

- **License**: Apache 2.0

[Open on Hugging Face ↗](https://huggingface.co/DavidAU/Llama-3.2-8X3B-MOE-Dark-Champion-Instruct-uncensored-abliterated-18.4B-GGUF)

DavidAU

### Dark Champion V2 8x4B

A larger 21B-class follow-up with a different expert composition.

- **License**: Apache 2.0

[Open on Hugging Face ↗](https://huggingface.co/DavidAU/Llama-3.2-8X4B-MOE-V2-Dark-Champion-Instruct-uncensored-abliterated-21B-GGUF)

Where to use it

## Hosted, community, and self-managed paths

01

### Managed hosted alternative

This exact family is not a public Vynaris endpoint. Use Qwen3.8 27B Uncensored when managed API access matters more than an exact checkpoint.

[View the hosted alternative →](https://vynaris.com/uncensored-models/qwen3-8-27b-uncensored)

02

### Hugging Face repositories

Inspect model cards, licenses, files, release notes, and community discussions directly from the publishers listed above.

[Open the first listed variant ↗](https://huggingface.co/DavidAU/Llama-3.2-8X3B-MOE-Dark-Champion-Instruct-uncensored-abliterated-18.4B-GGUF)

03

### Local or private deployment

Follow the selected repository's documented runtime path. Hardware, context, batching, and compatibility depend on that exact distribution.

[Read Vynaris API documentation →](https://vynaris.com/docs)

FAQ

## Questions about Llama 3.2 MoE Uncensored

**Is Llama 3.2 MoE Uncensored an official Meta model?**

No. The releases described here are independent community constructions based on Llama-derived components, not official Meta model variants.

**What does 8x3B or 8x4B mean?**

It describes the community model's expert composition. It does not mean every expert is fully active for every token, and it should not be used alone to estimate memory or speed.

**Can I access these models through an API?**

The linked repositories support self-managed use according to their instructions. Vynaris does not currently expose these exact builds, but it offers a related hosted dense 27B reduced-refusal model.

**Which Llama MoE variant should I choose?**

Compare the publisher's evaluations, license, runtime support, memory needs, and your own task-specific tests. The larger label does not guarantee better results for every workload.

Related model cards

## Compare nearby size and architecture classes

[Guide **GPT-OSS 20B Uncensored** 20B class](https://vynaris.com/uncensored-models/gpt-oss-20b-uncensored) [Guide **SuperGemma4 26B Uncensored** 26B class](https://vynaris.com/uncensored-models/supergemma4-26b-uncensored) [Hosted **Qwen3.8 27B Uncensored** 27B class](https://vynaris.com/uncensored-models/qwen3-8-27b-uncensored)

**Security-friendly, not criminal-friendly.**

Use these models only for lawful, authorized work. Child sexual exploitation, human trafficking or exploitation, and non-consensual sexual content are prohibited. Reduced refusal is not a safety, accuracy, or authorization guarantee.