Samsung Patents a Way to Blend Specialized AI Models Into One Without Starting Over
Training a powerful AI model from scratch is expensive and slow. Samsung's new patent describes a shortcut: take several already-trained models, blend their parts together with carefully tuned ratios, and end up with a single model that outperforms any of them individually.
How Samsung combines AI models that see and read
Today, AI teams often train separate models for separate jobs. One model might be great at reading text in photos; another might be better at answering questions about what's in an image. Getting the benefits of both usually means picking one and giving up the other, or spending a fortune training an entirely new model.
Samsung's patent describes a process for merging those models at a component level. Think of each model as having three sections: one that processes the image, one that handles language, and one that connects the two. Rather than fusing the whole model at once, Samsung's approach assigns a different mixing ratio to each section, then tests the result and adjusts until the combined model performs well across all the tasks.
The system keeps comparing candidate blends against each other automatically, picking whichever version scores higher on a defined goal. You'd never see this process directly, but it's the kind of work that determines how well your phone or tablet handles tasks like describing a photo, reading a document, or answering a question about something on screen.
multiplying the plurality of components with the first set of weights; and combining the multiplied plurality of components; …
Translation: It blends models by scaling and mixing their internal layers together.
How the component-wise merging and weight search works
A vision-language model (VLM) is an AI that can process both images and text at the same time, letting a device answer questions like "What does this sign say?" or "Who is in this photo?" Building a good one is costly, so researchers often fine-tune a base model for specific tasks rather than training fresh ones.
Samsung's method takes several of these models, including at least one general-purpose base and multiple fine-tuned variants, and merges them by treating each model as a set of three distinct components:
- A vision encoder (the part that interprets image content)
- A text encoder (the part that interprets language)
- A projection layer (the bridge that connects visual and language representations)
Instead of blending every layer the same way, the system assigns a separate numerical weight to each component and multiplies the components by those weights before combining them. Two candidate merged models are generated with different weight sets, each is tested against an optimization objective (a defined benchmark or quality score), and the better-performing one wins. The process iterates, steadily narrowing in on an optimal blend.
The core insight is that a text-heavy task might need more weight on the language component, while a visual task needs more on the vision side. Treating components separately lets the merge reflect those differences rather than applying one-size-fits-all averaging.
What this means for AI on Samsung devices
For people using Samsung devices with on-device AI features, this kind of work sits behind the scenes but shapes what those features can actually do. A phone that can accurately describe a photo, parse a receipt, or answer a question about a document is drawing on exactly the kind of multi-skilled model this patent is designed to produce. Samsung's steady investment in on-device AI points toward getting more of this capability running locally, without a server trip.
The practical payoff is that Samsung could ship a single model that handles a wide range of image-plus-language tasks, rather than maintaining multiple narrow models that eat into storage and processing budgets. For you, that likely means faster, more capable AI features that work even when your connection is poor.
This is the sixth Samsung filing we've tracked since July on AI models working in teams, after one where AI auditions answers and one promoting its top model.
When your phone's AI assistant describes a photo you took, it is drawing on separate learned skills, one for reading images and one for forming words, and those skills do not always cooperate. This patent describes a method for blending specialized AI models together more carefully, tuning how much each piece contributes so the result works better as a whole.
The practical payoff shows up in the small failures that stop happening. A caption that actually matches what is in the picture. An answer that reflects what the AI can clearly see rather than contradicting it.
Because this process runs on the device itself, without needing a server call, that reliability travels with you and does not depend on a strong connection. The improvement is quiet, but the frustrations it removes are real ones.
There are more where this came from
We read every patent application Big Tech publishes and send you the ones worth knowing. Plain English, free, every week.
The drawings
6 drawing sheets from US 2026/0268163 A1 · click any drawing to enlarge
Want this weekly breakdown for a company we don't cover? Patentlyze Pro →