Dolphin Mistral: what the Dolphin models are and how they're trained
What Dolphin Mistral is, who makes the Dolphin models, how they are fine-tuned to follow the system prompt instead of refusing, and their licenses.
By Unrestricted team · Editorial standards
Dolphin Mistral means the Dolphin fine-tunes that Eric Hartford and Cognitive Computations built on Mistral AI's open-weights models, such as Mistral 7B (Dolphin 2.x) and Mistral Small 24B (Dolphin 3.0).[1][2] Dolphin 2.x was trained on data with refusals and alignment-style answers filtered out, making it more compliant, and the 3.0 card says the system prompt sets the alignment.[1][2][3]
The base models are Mistral AI's; the fine-tuning is Dolphin's. Filtering the training data is one way to make an uncensored model; abliteration is another, and our guide to uncensored LLMs and how they're made covers both. This post covers who makes Dolphin, how it is trained, the Mistral-based versions and their licenses, using the public model cards, Hartford's own posts and the papers and pages cited below.
Who makes Dolphin and why
Eric Hartford announced Dolphin in July 2023 as an open-source, uncensored dataset and series of instruct-tuned models based on Microsoft's Orca paper.[3] The Dolphin 3.0 cards credit Eric Hartford, Ben Gitter, BlouseJury and Cognitive Computations, and say a grant from Andreessen Horowitz originally launched the project.[2][4]Earlier credits differ: the 2.2.1 card says a16z (Andreessen Horowitz) sponsored that model's training,[1] while the July 2023 announcement thanked EdenCoder and chirper.ai for mentorship and financial sponsorship.[3] The model cards now live under the dphn organization on Hugging Face.[2]
Hartford set out his reasons in a May 2023 essay, "Uncensored Models."[5]He wrote that open instruct models such as Alpaca, Vicuna and WizardLM were aligned because they were trained on data generated by ChatGPT. He argued that one alignment can't suit every culture and group, that it gets in the way of legitimate uses such as fiction with villains in it, and that building what he calls composable alignment has to start from an unaligned instruct model. On ownership he was blunt: "It's my computer, it should do what I want."[5]
The Dolphin 3.0 card makes a business version of the case. It says hosted assistants such as ChatGPT, Claude and Gemini control the system prompt, the model versions and a one-size-fits-all alignment, and can see your queries, while Dolphin "gives control to the system owner."[2] Both texts put the responsibility on the person using it: the essay says you are responsible for what you do with the output,[5]and the card calls you "the creator and originator" of anything you generate with it.[2]
How Dolphin is trained: dataset filtering and the system prompt
Dolphin began as an open rebuild of the Orca recipe. Microsoft's Orca paper (Mukherjee et al., June 2023) trained a 13-billion-parameter model on rich signals from GPT-4, including explanation traces and step-by-step thought processes, guided by teacher assistance from ChatGPT.[6]Hartford's team built a similar dataset: about 1 million FLANv2 prompts with GPT-4 completions and about 3.5 million with GPT-3.5 completions (the GPT-3.5 set after duplicates were removed). Then they filtered out what the announcement calls instances of alignment, refusal, avoidance and bias.[3]
Putting the essay, the announcement and the cards together, the method looks like this:[1][2][3][5]
- 01
Build the instruction data
Pair prompts with answers from larger models. The Dolphin announcement says the first dataset's answers came from GPT-4 and GPT-3.5.
- 02
Filter out refusals
Hartford's essay describes removing as many refusals and biased answers as possible and keeping the rest. The announcement lists alignment, refusal, avoidance and bias as what was filtered.
- 03
Fine-tune a base model on what's left
The essay says to train on the filtered data in exactly the same way the original model was trained. Five of the six Dolphin 2.x Mistral models in the table below start from Mistral 7B; the sixth, a 2.9.3 release, starts from Mistral NeMo 12B.
- 04
Leave the alignment to whoever deploys it
The announcement describes an uncensored model "upon which can be layered your personalized alignment LoRA" (a small add-on of fine-tuned weights). The Dolphin 3.0 card says the system prompt is "what you use to set the tone and alignment of the responses."
Each Dolphin 2.x card in the table below carries the same notice: the dataset was filtered to remove alignment and bias, which makes the model more compliant, and you are advised to add your own alignment layer before offering it as a service.[7][1][8][9][10][11] Later versions changed the data mix. Dolphin 2.2 added curated Samantha and WizardLM data for conversation and empathy,[1] 2.6 was trained with a lot of coding data,[8] and 2.9.3 supports function calling.[10]
The cards are also candid about the method's limits. Version 2.2.1 was a checkpoint release to fix overfit training: the earlier model answered with chain-of-thought reasoning when not asked, and it "was too compliant even when the request made no sense." [1]The 2.6 card says that because the model isn't DPO tuned, "you still might need to encourage it in the system prompt." [8]
Dolphin 3.0 added a new set of open datasets, including AllenAI's Tulu 3 SFT mixture, Hugging Face's SmolTalk, OpenCoder's SFT data, Orca AgentInstruct, NuminaMath and Hermes function calling, alongside Cognitive Computations' own Samantha and dolphin-coder data.[2]Its card credits a reward model used to filter the datasets and DeepSeek-V3 for augmenting the data, but it doesn't say what the filter removed. Unlike the 2.x cards, its text doesn't call the model uncensored (it links Hartford's essay instead), and it frames the model around steerability: "You set the system prompt. You decide the alignment."[2] The R1 version adds reasoning, trained for 3 epochs on 800,000 reasoning traces from the Dolphin-R1 dataset.[4]
The Mistral-based Dolphin models
This table covers Dolphin 2.x and 3.0 releases on Mistral AI base models, taken from their public cards; it isn't a full list of every Dolphin build. Where a card doesn't state something, the table says so rather than filling the gap.
| Model | Base model | Size | Context (per card) | License (per card) | Posted |
|---|---|---|---|---|---|
| Dolphin 2.0 Mistral 7B[7] | Mistral 7B | 7B | Not stated | Apache 2.0 | October 2023 |
| Dolphin 2.2.1 Mistral 7B[1] | Mistral-7B-v0.1 | 7B | Not stated | Apache 2.0 | October 2023 |
| Dolphin 2.6 Mistral 7B[8] | Mistral 7B | 7B | 16k base per card (Mistral's paper lists 8k)[12] | Apache 2.0 | December 2023 |
| Dolphin 2.8 Mistral 7B v0.2[9] | Mistral-7B-v0.2 | 7B | 32k base, fine-tuned at 16k | Apache 2.0 | March 2024 |
| Dolphin 2.9.3 Mistral 7B v0.3 32k[10] | Mistral-7B-v0.3 | 7B | 32k base, fine-tuned at 8k | Apache 2.0 | June 2024 |
| Dolphin 2.9.3 Mistral NeMo 12B[11] | Mistral-Nemo-Base-2407 | 12B | 128k base, fine-tuned at 8k | Apache 2.0 | July 2024 |
| Dolphin 3.0 Mistral 24B[2] | Mistral-Small-24B-Base-2501 | 24B | Not stated (the base card says 32k)[13] | Not stated | February 2025 |
| Dolphin 3.0 R1 Mistral 24B[4] | Mistral-Small-24B-Base-2501 | 24B | Not stated (the base card says 32k)[13] | Not stated | February 2025 |
Dolphin also shipped versions on Mistral AI's Mixtral models, including Dolphin 2.5 Mixtral 8x7b.[14]In Ollama's library, dolphin-mistral is the 7B line, with v2.8 as the default[15] and tags for v2, v2.1, v2.2, v2.2.1, v2.6 (plus a v2.6 DPO-laser build) and v2.8.[16]
Licenses. Mistral AI released Mistral 7B under Apache 2.0[17] and both the pretrained and instruction-tuned checkpoints of Mistral Small 3, its 24B model, under Apache 2.0 too.[18] The Dolphin 2.x cards in the table state Apache 2.0. The 2.0 and 2.2.1 cards call the model suitable for commercial or non-commercial use, and the 2.8 and 2.9.3 cards grant permission for any use, including commercial.[7][1][9][10][11]The two Dolphin 3.0 cards don't list a license;[2][4] when Hartford announced Dolphin, he wrote that each Dolphin model would be subject to the license of its base model.[3]Apache 2.0 grants a perpetual, royalty-free license to copy, modify and redistribute the work, with no clause limiting commercial use, as long as redistributions include a copy of the license, mark modified files, and keep the original copyright, patent, trademark and attribution notices, including any NOTICE file. The license doesn't grant trademark rights.[19]This is information, not legal advice: check the card and the base model's license before you ship anything built on them.
Dolphin vs abliterated models
Dolphin changes what the model learns: the 2.x models were fine-tuned on data filtered "to remove alignment and bias,"[1]starting from a dataset that dropped "instances of alignment, refusal, avoidance, and bias."[3]Abliteration instead edits an already-trained chat model: in 13 open-source chat models up to 72B parameters, Arditi and colleagues found that refusal is mediated by a single direction in the model's activations, and that erasing it stops the model refusing harmful instructions.[20] Our explainer on how abliteration removes refusals covers where the name comes from, the method and its side effects.
Running Dolphin locally
Ollama's dolphin-mistral tags are 4.1 GB at the default quantization and 14 GB unquantized,[16] the Dolphin 3.0 Mistral 24B card lists a 13 GB build for Ollama,[2]and Ollama's own dolphin3 entry is the Llama 3.1 8B version, not the Mistral one.[21] For hardware, runners and local vs hosted, see how to run an LLM locally, and for what else is in Ollama's library, our guide to Ollama uncensored models.
Using a Dolphin-Mistral-based model without running it
If you want a Dolphin-Mistral-based model without downloading weights, Unrestricted gives you a hosted model based on Dolphin-Mistral-24B, served by a third-party inference provider, on the web and iPhone. We didn't train it, and "based on" means just that: we don't claim it is identical to any release in the table above.
Saved chats are stored on our servers. On the web, incognito (new text-only chats) isn't saved to our database, but the prompt still goes to the model provider, as every prompt does. We add no filter or classifier of our own, but the model provider still refuses some prompts, and our Termsban some uses: violence planning, sexual content involving minors or non-consensual depiction of real people, targeted malware, fraud impersonation and illegal use. We don't train models on your chats.
Which model answers, and what it still won't do, is in the Unrestricted FAQ. If you're still choosing between fine-tuned, abliterated, local and hosted models, start with what uncensored LLMs are and how they're made.
Frequently asked
Is Dolphin Mistral uncensored?
Its makers describe the Mistral-based Dolphin 2.x models as uncensored: the dataset was filtered to remove alignment and bias, which the cards say makes the model more compliant. The Dolphin 3.0 card frames the idea as steerability, saying the system prompt sets the alignment. Uncensored describes how the model was trained, not a promise that it answers everything, and the 2.x cards advise adding your own alignment layer before offering it as a service.
Who makes Dolphin?
Dolphin was started by Eric Hartford, who announced it in July 2023. The Dolphin 3.0 cards credit Eric Hartford, Ben Gitter, BlouseJury and Cognitive Computations, and the models are published on Hugging Face under the dphn organization. For Dolphin Mistral, the base models come from Mistral AI and Dolphin does the fine-tuning.
What is Dolphin LLM?
Dolphin is a family of open-weights instruct models fine-tuned on filtered instruction data, so they are more compliant and are steered by the system prompt. It began as an open rebuild of the dataset behind Microsoft's Orca paper and has been released on several base models, including Mistral 7B, Mistral NeMo 12B, Mixtral, Mistral Small 24B and Meta's Llama 3.1 8B. Dolphin Mistral refers to the versions built on Mistral AI's models.
What license is Dolphin Mistral under?
The Dolphin 2.x Mistral cards state Apache 2.0; the 2.0 and 2.2.1 cards call the model suitable for commercial or non-commercial use, and the 2.8 and 2.9.3 cards grant permission for any use, including commercial. The Dolphin 3.0 Mistral 24B cards don't list a license, and their base model, Mistral Small 3, is Apache 2.0. Check the card and the base model's license before you redistribute or ship anything; this is information, not legal advice.
How big is the Dolphin Mistral download?
In Ollama's library, the dolphin-mistral 7B tags are 4.1 GB at the default quantization and 14 GB unquantized. The Dolphin 3.0 Mistral 24B card lists a 13 GB build for Ollama. How much memory you need depends on the quantization and the runner you use.
Sources
Checked November 2026.
- Dolphin 2.2.1 Mistral 7B model card, dphn on Hugging Face
- Dolphin 3.0 Mistral 24B model card, dphn on Hugging Face
- Dolphin (announcement, July 2023), Eric Hartford
- Dolphin 3.0 R1 Mistral 24B model card, dphn on Hugging Face
- Uncensored Models (May 2023), Eric Hartford
- Orca: Progressive Learning from Complex Explanation Traces of GPT-4 (Mukherjee et al., 2023), arXiv
- Dolphin 2.0 Mistral 7B model card, dphn on Hugging Face
- Dolphin 2.6 Mistral 7B model card, dphn on Hugging Face
- Dolphin 2.8 Mistral 7B v0.2 model card, dphn on Hugging Face
- Dolphin 2.9.3 Mistral 7B v0.3 32k model card, dphn on Hugging Face
- Dolphin 2.9.3 Mistral NeMo 12B model card, dphn on Hugging Face
- Mistral 7B (Jiang et al., 2023), arXiv
- Mistral-Small-24B-Base-2501 model card, Mistral AI on Hugging Face
- Dolphin 2.5 Mixtral 8x7b model card, dphn on Hugging Face
- dolphin-mistral, Ollama library
- dolphin-mistral tags, Ollama library
- Mistral 7B (announcement, September 2023), Mistral AI
- Mistral Small 3 (announcement, January 2025), Mistral AI
- Apache License, Version 2.0, Apache Software Foundation
- Refusal in Language Models Is Mediated by a Single Direction (Arditi et al., 2024), arXiv
- dolphin3, Ollama library
Keep reading
AI limits: message caps and rate limits in ChatGPT, Claude and Gemini
How AI usage limits work: message caps, rate limits and context windows in ChatGPT, Claude and Gemini, what 'unlimited' means, and how to tell which you hit.
ChatGPT Temporary Chat: what it saves, what it skips, and how long OpenAI keeps it
ChatGPT Temporary Chat skips your history, memory and training, but OpenAI can keep a copy for a while. How to start one, and how Claude and Gemini compare.
Abliterated models: what abliteration is and how it removes refusals
What an abliterated model is, how abliteration removes a model's refusal direction without retraining, its reported side effects, and what 'derestricted' means.