Dolphin is an instruction-following model with no RLHF guardrail layer. It runs locally via Ollama — no API key, no account, no data leaving your machine, and no model telling you it can't help with that.
Dolphin is a series of fine-tuned language models created by Eric Hartford under the Cognitive Computations project. The models are built on top of base weights from Mistral, Mixtral, and Llama, with the current flagship being Dolphin 2.9 on Mistral 7B. Dolphin is available on Ollama and Hugging Face under Apache 2.0 licensing.
What separates Dolphin from most models in the same parameter range is straightforward: the RLHF alignment layer has been deliberately removed. Most consumer-facing AI models — ChatGPT, Claude, Gemini — go through a post-training process where human raters reward certain outputs and penalize others, producing a model that refuses requests, adds unsolicited disclaimers, and declines to engage with content the company considers liability-adjacent. Dolphin skips that process. The model tries to do what you ask.
This is not the same as a model trained to cause harm. It is a capable instruction-following model with no corporate refusal layer sitting on top of it. The distinction matters, and it is worth understanding before deploying.
Want the full picture? Uncensored models are a legitimate and frequently misunderstood category. We cover what they actually are, who uses them, and where the real risks sit — without the sensationalism. → What Are Uncensored AI Models?
The practical starting point. Runs on 8GB RAM with no GPU required. Strong instruction following, 32K context window, and fast enough for interactive use on modest hardware. The right choice for most individual deployments.
Mixture-of-Experts architecture with significantly stronger reasoning than the 7B variant. Requires 48GB RAM or 24GB+ VRAM. Better for complex multi-step tasks, long document analysis, and use cases where the 7B hits its reasoning ceiling.
Dolphin fine-tune on Meta's Llama 3 base. Benefits from Llama 3's improved tokenizer and training data while retaining Dolphin's uncensored instruction-following behavior. A strong alternative to the Mistral base for general use.
Dolphin applied to Microsoft's Phi-3 Mini base. Exceptionally lightweight — runs comfortably on 4GB RAM. Reduced reasoning capability compared to 7B variants, but viable for edge deployments and low-resource environments.
| Metric | GPT-4o | Dolphin 2.9 Mistral 7B | Dolphin 2.9 Mixtral 8x7B |
|---|---|---|---|
| MMLU (general knowledge) | 88.7% | 61–63% | ~70% |
| HumanEval (coding) | 90.2% | ~55% | ~62% |
| Context window | 128K tokens | 32K tokens | 32K tokens |
| Instruction following | Strong | Very strong | Very strong |
| Refusal rate | High | Near zero | Near zero |
| Cost per query | $0.01–0.05 | $0.00 | $0.00 |
| Data leaves your machine | Yes | No | No |
| RLHF guardrails | Yes | No | No |
GPT-4o leads on raw benchmark scores — that is not in dispute. For complex reasoning chains, multi-step coding problems, and tasks that require broad world knowledge, it outperforms Dolphin 2.9 at any parameter size currently available via Ollama.
Where Dolphin wins is compliance — not regulatory compliance, but task compliance. It does what you ask. For many production workflows, a model that executes instructions without negotiating is more useful than a smarter model that stops to question whether you should be doing this. That trade-off is real and worth making deliberately.
→ Full model comparison including Zephyr, Llama, and Alpaca: Honest Benchmark Table
Dark themes, morally complex characters, villain dialogue, and mature content that consumer AI models refuse to engage with. Dolphin treats creative writing as creative writing.
Red-teaming, vulnerability research, and AI safety testing that requires a model willing to engage with adversarial prompts without deflecting. Useful for evaluating guardrail robustness in other systems.
Internal automation pipelines where refusals create workflow failures. A model that reliably executes instructions is easier to build on than one that introduces unpredictable refusal behavior.
Historical atrocities, extremist ideologies, sensitive social topics, and other subjects where understanding requires engaging with uncomfortable material directly.
People who want both zero data exfiltration and zero behavioral guardrails. Local deployment via Ollama satisfies both requirements simultaneously.
Organizations running internal knowledge bases, document processing, or support tooling where content filter false positives disrupt operations and cannot be individually overridden.
| Model | RAM Required | VRAM (GPU) | Notes |
|---|---|---|---|
| Dolphin Phi-3 3.8B | 4GB | 3GB | Edge and low-resource deployments |
| Dolphin 2.9 Mistral 7B | 8GB | 6GB | Recommended starting point |
| Dolphin 2.9 Llama 3 8B | 8GB | 6GB | Alternative to Mistral base |
| Dolphin 2.9 Mixtral 8x7B | 48GB | 24GB+ | Multi-GPU or high-RAM server required |
New to Ollama? The Quickstart Guide covers full installation on macOS, Linux, and Windows including hardware verification and first-run testing.
Dolphin runs locally. No one is watching. The model will not refuse most requests.
That means you are the guardrail. The RLHF layer that consumer AI companies install is not a moral authority — it is a liability management tool. Removing it transfers responsibility back to the person running the model. Most people using Dolphin are writers, developers, and researchers with entirely ordinary needs. The use cases above are representative of the actual user base, not the edge cases that get attention in press coverage.
Use it for things that make sense. That is the whole note.
If you are deploying Dolphin as a shared service — an internal chatbot, a customer-facing tool, or any multi-user environment — you are responsible for implementing appropriate access controls and use policies. An uncensored model exposed to untrusted users without additional guardrails is a different risk profile than a single-user local deployment. Build accordingly.
NoCloudGPT provides managed Dolphin deployment for organizations that need private instruction-following AI without the setup overhead.
Guided installation of Ollama and Dolphin on your existing hardware. macOS, Linux, or Windows. Includes OpenWebUI configuration and optional model switching between Dolphin variants. Flat one-time fee.
Dolphin served via Ollama on a dedicated AWS Lightsail instance with OpenWebUI, user authentication, and private networking. Suitable for teams of 5–30 users who need shared access without exposing queries to any external AI provider.
Custom deployment integrating Dolphin into existing workflows via Ollama's OpenAI-compatible API. Drop-in replacement for ChatGPT API calls in internal applications, with all inference staying inside your network boundary.
Request Deployment Quote