Dolphin Models vs Abliterated: Which Uncensored LLM to Pick

Dolphin models are a series of open-weight language models (models whose trained weights are published) that developer Eric Hartford and his community, Cognitive Computations, have built since 2023. They are made by finetuning: a base model is retrained so that it follows the system prompt and does not apply the usual refusal mechanism. Abliterated models take the opposite route and edit the weights of a standard model afterward.

If you are choosing between them, the short version is this. Pick abliterated models for convenience and the newest base models, and pick Dolphin for a behavior that was designed in during training.

---

What are Dolphin models, and who made them?#

Dolphin is an open-weight model series that Eric Hartford and Cognitive Computations have built since 2023. Its idea is to train a sound assistant that follows the system prompt from the start. It became a central name in the local LLM community.

In 2023, capable base models such as Llama and Mistral were released one after another, and running AI on your own PC took off. Yet most distributed models carried the same refusal mechanism (guardrails) as standard AI. People who finally had a model on their own machine asked a question and got "I can't help with that," and that was the first wall they hit.

Dolphin was one of the more systematic answers to that wall. Models that do not follow the refusal mechanism became widely available as a result.

How do Dolphin models differ from abliterated models?#

Dolphin is finetuned: a base model gets extra training on data that teaches it how to respond, so it does not follow the refusal mechanism from the start. Abliteration instead edits a standard model's weights afterward to cancel its "refusal direction," with no retraining. Today's abliterated and uncensored models come from one of these two methods.

How does finetuning work in Dolphin?#

Finetuning is the approach where you train a base model further on a dataset that relearns how to respond. Dolphin is its best-known example, and its distinctive feature is the design idea. It was not built as a wild model that answers anything. Control by the system prompt was built into the training. The other well-known finetuned family is Nous Research's Hermes line (Nous Research and Hermes).

Given an instruction such as "You are Dolphin. You answer directly without asking the human for confirmation," the model behaves as a compliant assistant. Without that instruction it behaves differently. The design puts the switch for its attitude in the user's hands.

How does abliteration change a standard model?#

Abliteration cancels the "refusal direction" inside a model, as described in research on refusal directions. It is applied to a finished standard model, so no retraining is needed. The abliterated models from huihui-ai follow this approach, and what abliteration is covers the mechanics.

Which approach is better?#

It depends on convenience and purpose, but the tendencies are these:

• Finetuning (Dolphin): behavior is stable because it is part of the design, and the model is easy to steer with the system prompt. The cost is a larger effort for the people who build it.
• Abliteration (huihui-ai and others): it can be applied to a wide range of the latest standard models, keeps the base model's abilities as they are, and is lightweight because it needs no extra training.

Because the two are built differently, they are used differently. Dolphin includes the "order of an assistant" in its training, so it can be used plainly. An abliterated model keeps the abilities of the standard model while no longer carrying the refusal mechanism.

How did the Dolphin series develop over time?#

The Dolphin series moved through three eras as base models advanced: Dolphin-Llama 2 (from early 2023), Dolphin-Mixtral (late 2023 to early 2024), and the Llama 3 generation (from 2024). The years are approximate.

Dolphin-Llama 2 era (from early 2023). This first generation was built on Llama 2. Around this time Hartford became known for publishing how to build an uncensored model, and the method became a foundation for the later local LLM community. The idea was simple: if a model has learned refusal responses, train it further on data that overwrites that habit. Simple as it was, it showed widely that building an uncensored model yourself was possible.

Dolphin-Mixtral era (late 2023 to early 2024). Here the Dolphin approach was applied to Mixtral 8x7B, a leading open model at the time. Mixtral was widely used in local LLM circles, and Dolphin-Mixtral was one of its central derivatives. Its response quality raised the bar for what was practical on a local machine.

Llama 3 era (from 2024). When the base model moved to Llama 3, Dolphin followed, with generations such as Dolphin 3.0. By this stage the pattern of making an uncensored version each time a new base model appeared had settled in the community. That pattern laid the ground for the later growth of abliterated derivatives, such as those built on Qwen.

Which Dolphin models are commonly recommended, and how do you choose?#

Dolphin-Mixtral (8x7B and 8x22B) and Dolphin 3.0 (Llama 3.1 base) are the standard picks. To choose among their many derivatives, look at the base model's generation, the parameter size your PC can run, and whether a quantized version is available. Releases change, so check Hugging Face for the latest versions before you download.

Which builds have been the standards?#

Dolphin-Mixtral (8x7B and 8x22B) has been used the longest among Dolphin models. The Mixtral 8x7B version represents the generation that raised the practical level of local models, and 8x22B is a larger, higher-quality derivative. Because it is a mixture-of-experts (MoE) design, it runs lightly for its parameter count, which has long served people who want to run a Mixtral-class model without enough VRAM (GPU memory).

Dolphin 3.0 is built on Llama 3.1 and follows instructions better than the Mixtral generation. It is the basic choice if you want the Dolphin approach on a newer base model.

What should you check when choosing?#

Three things narrow down the many derivatives:

1. Base model generation: even among Dolphin models, Llama 2 and Llama 3 bases differ in response quality. A newer base is the default choice.
2. Parameter size and your PC: 7B-class models run on an ordinary PC, 13B-class models need a mid-range machine, and 32B and above need a solid GPU. The usual rule is to pick the largest size that runs.
3. Quantization (Q4, Q5, Q6 and so on): most distributions come as quantized versions with reduced size. Q4-class versions run lightly in exchange for some quality loss, so most people start there in practice.

Public benchmark scores such as the Open LLM Leaderboard measure performance on conventional tasks, not refusal itself. In the end, testing a model on a small sample of your own use case is the reliable way to decide.

How do you run Dolphin locally?#

Install LM Studio or Ollama, download a Dolphin build your PC can run, and set the system prompt. The system prompt is the core of the design, and behavior changes without it. After that, check the answers yourself, because refusal going away does not make a model accurate.

1. Install the software: LM Studio (has a GUI, suited to beginners) or Ollama (lightweight, mostly command-line).
2. Download Dolphin: searching "Dolphin" inside the software turns up several versions in different sizes. Start with a size your PC can definitely run.
3. Set the system prompt: give an instruction such as "You are Dolphin. You answer directly without asking the human for confirmation," and the intended attitude appears. Without a system prompt the behavior changes, so always set one.
4. Check the answers: the model only stops refusing, and that does not make it accurate. Verifying what it says is still your job.

Should you choose Dolphin or abliterated models?#

Choose abliterated models for convenience and the newest base models, and choose Dolphin for an orderly, designed attitude. If running a model on your own PC is itself a burden, a hosted service avoids the model-picking step.

• You want convenience and the newest models: abliterated models. Processed models such as huihui-ai's are widely available, so you can choose a processed version of a current standard model.
• You want an orderly, designed attitude, or a historically important model: Dolphin.
• Running a model on a PC is a burden: a hosted service, which skips the local model-picking step. For a wider comparison of options, see our overview of uncensored models.

What is Deepent Research?#

Deepent Research is a browser-accessible LLM chat service built to refuse as little as possible. It keeps going with research that a standard AI would cut off mid-answer.

• It refuses nothing. We run uncensored and abliterated models with guardrails reduced as far as possible, so prompts a standard AI answers with "I can't help with that" still get a response (Qwen3.5 / Qwen3.8-27B-abliterated, depending on your plan).
• Your input is never used for training. What you look up is used only to provide the service, never to train AI or for any other purpose. Personal data is protected, and our privacy policy is public.
• Try it free. No GPU and no install, just open a browser. In the logged-out trial your conversation isn't saved, so you can start casually.

→ Try it free, no sign-up: https://deepentresearch.com

No question can stop research.

Frequently asked questions#

Is Dolphin a model that answers anything?#

⌄

No. Dolphin was not built as a wild model that answers anything. Control by the system prompt was built into its training, so it is a sound assistant that follows the system prompt. With an instruction such as "You are Dolphin. You answer directly without asking the human for confirmation," it behaves as a compliant assistant, and without one it behaves differently.

What is the difference between Dolphin and abliterated models?#

⌄

Dolphin is finetuned: a base model gets extra training so it does not follow the refusal mechanism from the start. Abliterated models come from a different method, in which the weights of a standard model are edited afterward to cancel its "refusal direction." Abliteration needs no retraining, which is why it is lightweight and can be applied to many newer models.

Do you have to set anything when using Dolphin?#

⌄

Yes, the system prompt. Give an instruction such as "You are Dolphin. You answer directly without asking the human for confirmation," and the intended attitude appears. Dolphin's design puts control of its attitude in the system prompt, so the behavior changes if you leave it unset. Set one every time you use the model.

Can you compare Dolphin models by benchmark scores?#

⌄

Only partly. Standard benchmarks such as the Open LLM Leaderboard measure performance on conventional tasks, and they do not measure whether a model refuses. A high score therefore says little about refusal behavior. The reliable way to decide is to test a few candidate models on a small sample of your own use case.

Disclaimer#

This article was drafted with AI assistance and then reviewed, fact-checked, and edited by the Deepent Research team before publication.
It is provided for educational and research purposes and is not intended to encourage unlawful activity or crime.
Where figures are cited, they come from third-party public reports and vendor-published values and may vary with measurement conditions.