Open-source models trained to sound like GPT-4o
Open models that anyone can download and run, whose makers say they were trained to sound like GPT-4o.
These are different models. None of them is GPT-4o. An open model is one whose files are public: you run it on your own computer, or on one you rent.
Every claim about style on this page is the maker's own. None of these models has a public evaluation against GPT-4o.
We asked each of them the same questions as GPT-4o , on our own computer. There is no result yet on how close any of them sounds.
Helcyon GPT-4o (Gemma 4 12B)
By XeyonAI
The maker says it was fine-tuned to recreate GPT-4o 's warmth, humor and conversational style, on GPT-4o data that it does not name.
Built on Gemma 4 12B, according to the maker's page
License Not stated
Download size 8.55 GB or 9.79 GB, by quality. A separate 175 MB file lets it read images.
Last updated November 3, 2026
We asked it the same questions as GPT-4o , on our own computer. Read its answers beside GPT-4o 's
Helcyon-4o v4.0 (Mistral Nemo 12B)
By XeyonAI
The maker says it carries GPT-4o 's tone and was compared with GPT-4o. No numbers are published.
Built on Mistral Nemo Instruct 2407
License Apache 2.0
Download size 8.73 GB or 10.06 GB, by quality.
Last updated May 23, 2026
We asked it the same questions as GPT-4o , on our own computer. Read its answers beside GPT-4o 's
Helcyon-GPT-4o v4.5 (Mistral Nemo 12B)
By XeyonAI
The maker says its training data was written by GPT-4o itself, and calls this version the closest in tone so far.
Built on Mistral Nemo Base 2407
License Apache 2.0
Download size 10.06 GB for the one version we checked.
Last updated July 26, 2026
We asked it the same questions as GPT-4o , on our own computer. Read its answers beside GPT-4o 's
Nova-Qwen-14B
By Lighstromo
The maker says it was fine-tuned on 3,066 GPT-4o conversations, taken from a cleaned copy of the set used by the two distillations below.
Built on Qwen2.5 14B Instruct
License Apache 2.0
Download size 8.99 GB or 12.12 GB, by quality.
Last updated March 2, 2026
We asked it the same questions as GPT-4o , on our own computer. Read its answers beside GPT-4o 's
GPT-4o distillation on Llama 3.1 8B
By trentmkelly (the add-on) and Meta (the base model)
An add-on trained on 12,286 examples to prefer the answers of ChatGPT's own GPT-4o over the base model's answers. Its maker suggests it as the place to start for a GPT-4o fan with an ordinary computer.
Built on Meta's Llama 3.1 8B Instruct
License Not stated for the add-on. Meta's Llama license covers the base model.
Download size The add-on is 168 MB. It is not a complete model: it is applied to Meta's base model, which you download separately.
Last updated January 30, 2026
We asked it the same questions as GPT-4o , on our own computer. Read its answers beside GPT-4o 's
GPT-4o distillation on Llama 3.3 70B
By trentmkelly (the add-on) and Meta (the base model)
The same recipe on a much larger base model. Its maker notes that most people cannot run it.
Built on Meta's Llama 3.3 70B Instruct
License Not checked by us for the add-on. Meta's Llama license covers the base model.
Download size The add-on is 829 MB, the size of our own copy. It is applied to Meta's base model, which you download separately.
Published February 2, 2026
We asked it the same questions as GPT-4o , on our own computer. Read its answers beside GPT-4o 's
Aura-4o-Rebirth (Gemma 4 31B)
By SevenOfNine, on Hugging Face
The maker says it was trained on 16,509 question-and-answer pairs rebuilt from one person's 2.7 years of GPT-4o conversations.
Built on Gemma 4 31B
License Apache 2.0
Download size 21.85 GB for the version we ran, the size of our own copy. The same page has two other versions, by quality. A separate 1.2 GB file lets it read images.
Maker's page dated May 4, 2026
We asked it the same questions as GPT-4o , on our own computer. Read its answers beside GPT-4o 's
File sizes, licenses and dates were read on Hugging Face on November 18, 2026, unless a line says otherwise.