DEV.co
Open-Source LLM · DavidAU

Llama-3.2-8X3B-MOE-Dark-Champion-Instruct-uncensored-abliterated-18.4B-GGUF

A Llama 3.2 mixture-of-experts model (18.4B parameters from 8×3B experts) optimized for creative writing, fiction, and roleplay. Runs at 50+ tokens/sec on low-end 16GB GPU. Uncensored output. Apache 2.0 licensed, not gated. Requires explicit sampling tuning (temperature, samplers) and selection of expert count (1–8) to control output quality and variety.

Source: HuggingFace — huggingface.co/DavidAU/Llama-3.2-8X3B-MOE-Dark-Champion-Instruct-uncensored-abliterated-18.4B-GGUF
Unknown
Parameters
apache-2.0
License (OSI-approved)
Unknown
Context (tokens)
39.7k
Downloads (30d)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

FieldValue
DeveloperDavidAU
ParametersUnknown
Context windowUnknown
Licenseapache-2.0 — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads39.7k
Likes578
Last updated2026-04-28
SourceDavidAU/Llama-3.2-8X3B-MOE-Dark-Champion-Instruct-uncensored-abliterated-18.4B-GGUF

What Llama-3.2-8X3B-MOE-Dark-Champion-Instruct-uncensored-abliterated-18.4B-GGUF is

MOE architecture combining 8 Llama 3.2 3B expert models (including abliterated and uncensored variants) via Mergekit. Default 2-expert gating; user-configurable 1–8 experts. Quantized to GGUF; multiple quants available (IQ4XS tested). 128k context. Operates at bfloat16 internally; stable across sampling parameters (temp 0–5). Different expert combinations yield different outputs per generation. Requires llama.cpp-compatible inference (LMStudio, KoboldCpp, text-generation-webui, llama-server).

Quickstart

Run Llama-3.2-8X3B-MOE-Dark-Champion-Instruct-uncensored-abliterated-18.4B-GGUF locally

Load the open weights with 🤗 Transformers and generate — the same model, self-hosted.

quickstart.pypython
from transformers import pipelinepipe = pipeline("text-generation", model="DavidAU/Llama-3.2-8X3B-MOE-Dark-Champion-Instruct-uncensored-abliterated-18.4B-GGUF")out = pipe("Explain retrieval-augmented generation in one sentence.",           max_new_tokens=128)print(out[0]["generated_text"])

Swap in vLLM or Ollama for production-grade serving. DEV.co can stand up the inference stack.

Deployment

How you'd run it

A typical self-hosted path — open weights, an inference server, your application.

DEV.co builds each layer — from GPU infrastructure to the application.

Best use cases

Fiction and Story Writing

Designed for creative writing across all genres. Model card claims exceptional output for prose, fiction, and storytelling, with configurable tone (e.g., 'vivid horror') via prompting and expert selection.

Roleplay and Interactive Fiction

Supports multi-expert routing for variety. Users can adjust expert count (2–8) and temperature to balance instruction-following (fewer experts, lower temp) with prose variety (more experts, higher temp).

Fast Local Inference on Consumer GPU

Quantized GGUF format achieves 50+ t/s on 16GB low-end GPU with IQ4XS, or 2× that on mid-range cards. Suitable for private, self-hosted LLM deployments where speed and VRAM efficiency matter.

Running & fine-tuning it

Estimated 8–12 GB VRAM for inference at IQ4XS quantization on low-end GPU (50+ t/s); 6–8 GB on mid-range GPU (100+ t/s estimated). Full model unquantized: ~18.4B parameters in bfloat16 ≈ 36 GB. Requires llama.cpp-compatible runtime (CPU inference possible but slow). Exact VRAM/precision trade-off requires verification with target inference framework.

Unknown. Model card does not discuss LoRA, QLoRA, or fine-tuning feasibility. Derived from base Llama 3.2 and merged via Mergekit; standard Llama 3.2 fine-tuning techniques may apply, but stability and expert routing behavior under gradient updates is not documented. Requires testing.

When to avoid it — and what to weigh

  • Requirement for Consistent, Deterministic Output — MOE gating selects different expert pairs per generation even for identical prompts. Model card notes 2–4 regenerations may be needed to achieve desired quality. Not suitable if reproducibility is critical.
  • Safety-Critical or Production System Requiring Content Moderation — Model is explicitly abliterated and uncensored. Model card warns of NSFW, visceral details, swearing, and light horror. Contains uncensored submodels that may interfere with safety guarantees.
  • General-Purpose Chat or Instruction-Following at Scale — While model supports general use, primary design and tuning is for creative writing/roleplay. Unknown parameter count and no general-purpose benchmarks provided. Base Llama 3.2 models may be more reliable for diverse tasks.
  • Production Deployment Without Hands-On Tuning — Model requires manual configuration of expert count, temperature, sampler choice (Dry, Dynamic Temp, Smooth), and quant level. No 'out-of-box' defaults stated for non-creative use cases.

License & commercial use

Apache 2.0 (Apache License 2.0). This is a permissive OSI-approved open-source license. Permits commercial use, modification, and distribution, provided the license and copyright notice are retained.

Apache 2.0 permits commercial use, modification, and redistribution. However, the model is a merge of 8 base models, including Meta Llama 3.2 (which requires acceptance of Meta's Llama 3.2 community license) and third-party uncensored/abliterated variants. Review the licenses and terms of all constituent models before commercial deployment. The Apache 2.0 license on this specific merge repo does not override upstream license obligations. Recommend legal review.

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

SignalAssessment
MaintenanceActive
DocumentationAdequate
License clarityNeeds review
Deployment complexityModerate
DEV.co fitGood
Assessment confidenceMedium
Security considerations

Model is explicitly uncensored and contains abliterated submodels designed to bypass safety training. Output may include NSFW, visceral, or harmful content. Suitable only for contexts where such content is acceptable (private creative writing). Deployment in customer-facing or moderated systems requires additional content filtering. No adversarial robustness, prompt injection, or jailbreak evaluations provided. Treat as high-risk for safety-sensitive applications.

Alternatives to consider

Meta Llama 3.2 3B / 8B / 1B (base models)

Official, safety-aligned alternatives with broader benchmarking and production support. Slower for creative writing but more predictable and moderation-friendly.

Mistral 7B Instruct (or variants)

Widely deployed, well-documented, similar parameter range, better general-purpose instruction-following. Less specialized for creative writing but more stable.

OpenAI GPT-4 or Claude 3.5 (API)

If consistency, safety, and multi-task capability are required and local inference is not mandatory. Higher cost but reduced tuning and deployment complexity.

Software development agency

Ship Llama-3.2-8X3B-MOE-Dark-Champion-Instruct-uncensored-abliterated-18.4B-GGUF with senior software developers

Evaluate this MOE model for private fiction, roleplay, and story generation. Test quantization, expert count, and sampling parameters on your hardware. For production safety requirements or multi-task deployment, consider safety-aligned alternatives. Contact us for custom fine-tuning or integration support.

Talk to DEV.co

Related open-source tools

Surfaced by semantic similarity across the DEV.co open-source index.

Related on DEV.co

Explore the category and the services that help you build with it.

Llama-3.2-8X3B-MOE-Dark-Champion-Instruct-uncensored-abliterated-18.4B-GGUF FAQ

Can I use this model for commercial applications?
Apache 2.0 permits commercial use of this repo. However, the model merges Meta Llama 3.2 and third-party uncensored models, each with their own license terms. Verify Llama 3.2 community license and the license of each constituent model before commercial deployment. Legal review recommended.
What GPU do I need to run this?
Quantized GGUF (IQ4XS) requires ~8–12 GB VRAM for 50+ t/s on low-end GPU. Mid-range GPUs achieve ~100+ t/s (estimate). Exact requirements depend on inference framework, quantization level, and batch size. CPU inference is supported but slow. Test with your target hardware and quant.
Why does the output differ every time I run the same prompt?
The model uses mixture-of-experts with configurable expert gating (default 2 out of 8 experts). Different expert combinations are selected per generation, causing output variation. This is by design for creative writing variety. Use fewer experts or lower temperature for more consistency.
Is this model safe for production chatbots?
No. The model is explicitly uncensored and abliterated. It can output NSFW, visceral, and unmoderated content. It is designed for private creative writing and roleplay, not customer-facing systems. Requires external content filtering for any public deployment.

Work with a software development agency

DEV.co is a software development agency delivering custom software development services to companies building on open source. Our software developers and web developers design, integrate, and ship production systems — spanning web development, APIs, AI, data, and cloud. If Llama-3.2-8X3B-MOE-Dark-Champion-Instruct-uncensored-abliterated-18.4B-GGUF is part of your open-source llms roadmap, our team can implement, customize, migrate, and maintain it.

Ready to Deploy a Creative-Writing LLM Locally?

Evaluate this MOE model for private fiction, roleplay, and story generation. Test quantization, expert count, and sampling parameters on your hardware. For production safety requirements or multi-task deployment, consider safety-aligned alternatives. Contact us for custom fine-tuning or integration support.