The other day I was watching a youtube video about models such as Fable and the host of podcast mentioned people from China, Russia or Iran may distil these models and make their own versions and called it dangerous.
Honestly, what I found out – as an Iranian individual – is how Dario Amodei (and in general people like him) made this type of propaganda to keep their best models a huge money-making machine and nothing more. Well, you can’t stop us from using models which are already hosted on OpenRouter. You can’t keep us from using Ollama or similar tools. Can you?
Well this is not the whole story. I was messing with Qwen 3.5 as soon as it got released and I found out how dynamic this model is in terms of context window size and also reasoning. But since internet connection was super expensive back then and our country was at war, I didn’t bother to work on the model that much.
Until “Fable” came out!
When this model came out, I saw how many people lost their crap over it. It was really a good model in any way possible and it could pass pretty much any test you had in mind.
After a few days, the model just went offline for “non American” people, even the ones living in US borders (imagine Andrej Karpathy can’t use a production of his own!). Well this made a huge desparation world wide.
GLM 5.2 became the next cool kid in town, but it still had a problem. It was too big to be self-hosted. So now what is the solution?
Maral Mythos 9B is!
Well, I had a simple secret dataset, and every time a new SOTA model was available on Open Router or Cloudflare AI (or any other aggregator like them) I have it updated.
The dataset, is around 500 million tokens of Claude 4.8 Opus and Claude 5 Fable data. It is data from GPT 5.5 and similar models. So it’s worth a try on Qwen 3.5, isn’t it?
And after the fine tune, the benchmark looks like this:
So, I have decided to release the model to the public because we need open source models to fight “personal decisions” which can affect lives of millions.
Links to the models
Original Model: https://huggingface.co/MaralGPT/MaralGPT-Mythos-9B-2606
GGUF Models: https://huggingface.co/MaralGPT/MaralGPT-Mythos-9B-2606-GGUF
If you can cooperate on further development or deployment of the model, you can stay in touch with me here or over haghiri75[at]gmail.com.