MaralGPT "Mythos" 9B just released, and this is why I'm proud of my project

A tiny underdog AI dropped, the site melted, and the comments went feral

TLDR: MaralGPT Mythos 9B was released as a small open AI alternative, with its creator pitching it as a pushback against big companies controlling access. The community split between cheering the underdog, roasting the site crash, and raising eyebrows at the model’s mysterious training recipe.

A new small AI model called MaralGPT Mythos 9B has arrived, and the creator isn’t being shy about the mission: this is being framed as a rebellion against powerful companies deciding who gets access and who gets locked out. The post comes with real emotion too, tying the project to frustration over blocked access, expensive internet, and the feeling that ordinary people outside the biggest tech circles are being told to sit down and stay out. That alone lit a fire under readers.

But the real popcorn moment was the community reaction. One of the first vibes was pure admiration — a simple “Wow Respect” set the tone for the fans. Then the drama instantly turned slapstick: people noticed the creator’s personal website was buckling under the attention, with commenters joking that getting to the top of Hacker News basically hugged the site to death. Within minutes, the crowd went full internet survival mode, racing to post Wayback Machine and archive links like emergency responders at a digital crime scene.

And then came the skeptical side-eye. One commenter boiled the whole thing down to a much spicier summary: this is basically a smaller model trained on a “simple secret dataset” made up of output from famous premium chatbots. That’s where the hot take lives: is this an open-source hero story, or just another remix with a rebellious press release? Either way, the comments made one thing clear: the launch wasn’t just a model drop — it was a full-blown main character moment.

Key Points

  • Muhammadreza Haghiri announced the public release of MaralGPT "Mythos" 9B on June 29, 2026.
  • The article says the model is based on Qwen 3.5, which the author had evaluated for its context-window flexibility and reasoning ability.
  • Haghiri states that the model was fine-tuned on a proprietary dataset of about 500 million tokens.
  • According to the post, the dataset contains outputs from Claude 4.8 Opus, Claude 5 Fable, GPT 5.5, and similar models accessed through services such as OpenRouter and Cloudflare AI.
  • The article provides Hugging Face links for the original model and GGUF versions and invites collaboration on further development or deployment.

Hottest takes

"Wow Respect" — vanessa1211
"Getting to the top of HN seems to have crashed this site" — wateralien
"simple secret dataset" — zihotki
Made with <3 by @siedrix and @shesho from CDMX. Powered by Forge&Hive.