When you're logging in to edit a project, you have a ton of footage to go through. It doesn't matter if it's short-form web content or a whole movie; there's always a bunch of angles and ideas to view. And not only do you want that to be nicely organized, but you also want it secure, so only the final product gets out into the world.
Eddie AI saw those problems across their clients and decided to do something about it. They announced progress on a specialized 9-billion-parameter model designed to learn privately and exclusively for individual customers.
The research done so far has shown the compact model achieving up to a 47% win rate against Kimi K3 when tasked with finding narrative arcs in raw footage.
Now, if you're an editor, you know this signals a future where you can get a faster rough cut or a personalized version of any project without losing control of it.
Let's dig a little deeper into this one.
- YouTube www.youtube.com
Specialization Beats Raw Scale
Are you ready for some technical specs? Eddie AI trained a smaller model (Qwen3.5-9B) by using reasoning traces from Kimi K3 to create the new 9B model.
Then, they tested it on 100 real projects, but made sure they focused on fast-paced social video edits. That way, they could get it to understand hooks and other concepts.
This allowed the new program to skyrocket in performance.
- Off-the-shelf base model: ~10% win rate against Kimi K3
- Initial LoRA fine-tuning: ~32% win rate
- Refined iterative training (best-of-N sampling & filtering): Up to a 47% win rate
“A general-purpose model has to be good at everything. A specialized model only has to become exceptionally good at editing video,” explained Alexander Terekhov, Co-founder and Head of AI at Eddie AI. “This result gives us confidence that smaller, customer-specific systems can close the gap on valuable editorial decisions while remaining practical to deploy.”
Complete Knowledge Ownership
Okay, so you have a ton of speed, but what about security?
Because the 9B model features a small footprint, it can be deployed on-premises.
The system learns a customer's specific pacing, B-roll preferences, and visuals to stay consistent with what has been edited.
“Customers are not opposed to AI training—they are opposed to losing control of the knowledge they create,” said Shamir Allibhai, Co-founder and CEO of Eddie AI. “Our goal is to give every customer a system that learns only with their permission, improves exclusively for them, and keeps their creative advantage theirs.”
Eddie AI plans to expand the model's capabilities beyond rough narrative cuts to encompass complex visual assembly, music placement, and motion graphics.
Summing It All Up
To get started with faster rough cuts while keeping your data private, try the core assistant today at www.heyeddie.ai.
We find it best to reach out through their contact sales page to learn more about private enterprise model deployments, or meet the team in person if you are heading to Amsterdam for IBC 2026.
Let us know what you think in the comments.
