What are uncensored LLMs?

Uncensored LLMs are open-weight language models that have been adjusted to mitigate the refusal mechanisms commonly found in standard AI assistants. By offering greater user control over model behaviour, they become particularly relevant for individuals who host and experiment with LLMs on local hardware.

What are uncensored LLMs?

The majority of contemporary AI assistants are designed to adhere to safety guidelines and decline specific types of requests. This behaviour typically stems from instruction tuning, preference training, system prompts, or various components of the model and application.

An uncensored LLM is, in most cases, a model that has been altered or trained to diminish these restrictive tendencies. There is no single technical definition for "uncensored". Different creators employ distinct methods, leading to models that exhibit significantly different behaviours.

Some uncensored models emerge from additional fine-tuning, while others utilise techniques that target specific behavioural traits within an existing model. The term may also apply to models described as abliterated; however, abliteration is a specific technical approach rather than a synonym for all uncensored models.

Uncensored does not mean unrestricted

Reducing or removing refusal behaviour does not inherently enhance a model's capability. An uncensored model may still generate inaccurate information, misinterpret instructions, or decline certain requests.

• Capability remains critical: A smaller model will not improve as a reasoner solely because its refusal behaviour has been modified.
• Quality is variable: The performance of uncensored models can vary widely depending on the base model and the specific modifications applied.
• Behaviour is not guaranteed: Even uncensored models may occasionally refuse requests or follow instructions inconsistently.
• Safety measures may shift: Reducing refusals can inadvertently remove safeguards that were integral to the original model's training.

Consequently, it is more prudent to view "uncensored" as a descriptor of the model's behavioural profile rather than a guarantee of its functional limits.

Uncensored vs open-weight vs base models

While these terms are frequently used in conjunction, they denote distinct characteristics of an LLM.

These categories often intersect. An uncensored model may be open-weight and derived from an existing base. It can also be a fine-tuned variant or another specific modification of that model. The label alone does not fully elucidate the methodology behind its creation.

Why run an uncensored LLM locally?

Hosting an uncensored LLM locally grants the user superior control over the model and its operational environment. Rather than relying on a hosted AI service, the model executes on hardware under the user's direct management.

• Control: You select the model, inference software, and configuration parameters.
• Privacy: Prompts and generated responses remain within your own computing environment.
• Customisation: Open-weight models can be modified, fine-tuned, and configured for diverse workloads.
• Offline use: A locally hosted model does not require prompts to be sent to an external AI service.
• Experimentation: Developers and researchers can evaluate different model versions and modifications.

Local inference also affords control over the hardware executing the model, a factor that becomes increasingly significant as model sizes expand.

What hardware do uncensored LLMs require?

Uncensored models generally possess the same hardware requirements as the base models upon which they are built. Key factors include model size, quantisation, context length, and inference settings.

Larger models demand more memory than smaller counterparts. Quantisation can lower the memory required to load a model, rendering larger architectures practical on GPUs with limited VRAM.

VRAM is also consumed by the inference process itself. The KV cache and other runtime data necessitate additional memory, and extended context windows can further increase memory demands.

Therefore, selecting a model is only one aspect of planning a local LLM setup. The GPU must possess sufficient available VRAM to support both the model and the intended workload.

Try on DaDesktop

If you wish to run an uncensored LLM without purchasing and installing dedicated GPU hardware, DaDesktop offers cloud desktops equipped with dedicated GPU resources. You can run local LLM workloads on DaDesktop or compare available GPU options tailored to your chosen model.