@artyomita's reel — transcript & breakdown

Starts by showing a complex, long model name and highlighting its high download count to establish credibility and interest.

“This is a real Qwen model name. 1.7 million downloads last month.”

Explains 'Abliterated' with a visual representation of weights being removed, making a complex concept easy to understand.

“Let's read it like a label. Abliterated: someone found the direction in the weights that says no, and deleted it. The safety goes with it.”

Continues explaining 'Heretic' with an automatic removal visual, demonstrating efficiency and automation.

“Heretic: a free tool that does the same thing, automatically.”

Describes 'Distilled' using a student-teacher analogy and visual comparison of model sizes, simplifying technical jargon.

“Distilled: a smaller model trained on a bigger model's answers. A student copying the teacher.”

Introduces 'LoRA' as an add-on, showing its physical attachment to weights and giving a practical size example.

“LoRA: a small add-on that clips onto the weights. I make these: one of mine is 200 megabytes.”

Explains 'MTP' as extra heads for faster predictions, with a visual that abstracts how it works and states a personal contribution.

“MTP: extra heads that guess the next few words, so it runs faster with the same answers. I published MTP heads for Qwen3.8 Flash-Next: 1595 downloads last month.”

Summarizes the different manipulations possible on a single model, highlighting versatility.

“Same model. Five different things done to its weights.”

Direct call to action to save the information and directs viewers to the creator's GitHub for more work.

“Save this before your next download. My work is on my GitHub.”