Eddie AI Unveils a Specialized 9B Model Built for Private, Personalized Video Editing
Eddie AI has announced a 9-billion-parameter model designed specifically for video editing workflows, with a focus on private, per-customer deployment. Rather than relying on a single large shared model, the system is trained to learn each customer's editorial preferences - pacing choices, B-roll habits, visual style - and improve exclusively for that client. The announcement comes alongside early benchmark results that suggest the approach has meaningful technical legs.
The model was built by fine-tuning Qwen3.5-9B using reasoning traces distilled from Kimi K3, a much larger general-purpose system. Eddie AI tested the model across 100 real projects, with particular attention to fast-paced social video formats that require an understanding of hooks and narrative structure. Starting from a base win rate of roughly 10% against Kimi K3, the team applied iterative LoRA fine-tuning and best-of-N sampling to push performance to a 47% win rate - close to the 50% that would represent full parity. Notably, the final model's reasoning traces were 30% shorter than earlier versions, making it practical to run on standard GPUs rather than expensive inference clusters.
The results so far are scoped to transcript-level story construction - identifying narrative arcs within raw footage - rather than full visual assembly. But they demonstrate a broader principle: a smaller model trained on a narrow, well-defined task can close much of the gap on larger general-purpose systems. For editors, the near-term practical value is faster rough cuts with a system that already understands how a given team or studio tends to work, without requiring manual re-briefing on every project.
On the privacy side, the small footprint of the 9B model is what makes on-premises deployment viable. Customers retain full ownership of the training data and the knowledge the model develops from it - nothing feeds back into a shared system. Eddie AI has stated plans to extend the model's capabilities beyond narrative structuring to include visual assembly, music placement, and motion graphics. The team is expected to be present at IBC 2026 in Amsterdam for anyone looking to discuss enterprise deployment options in person.
