Merve Noyan

Noyan’s work combines implementation with teaching. Hugging Face introduced her as a developer advocate building tools and content to make machine learning accessible. Her public projects focus on zero-shot vision, multimodality and optimization. Her smol-vision repository collects notebooks and recipes for shrinking, optimizing and customizing visual models, giving developers examples they can inspect and adapt.[1][4][6]

Small models that can understand images

Noyan coauthored the SmolVLM introduction. The release presents a two-billion-parameter vision-language model family for smaller local setups, with base, synthetic-data-tuned and instruction variants. It publishes checkpoints, datasets, training recipes and fine-tuning tools. The design treats memory use and deployability as central constraints, so multimodal capabilities can be explored without assuming access to the largest machines.[2]

Let an agent use what it sees

Her coauthored guide to visual agents explains how smolagents can put images into an agent’s memory. A callback records a visual observation after a tool action, allowing the next model step to reason about the resulting state. The browser example combines navigation tools with screenshots. Its concrete loop connects the agent’s action to evidence of what happened, which is essential when a website’s state changes.[3]

In her conference talk, Noyan presents a workflow in which vision-language models help label data and evaluate annotations before training a specialized detector. The system includes human review of prompts and generated labels. The distinction matters for deployment: a flexible general model can help build a task-specific computer-vision model whose runtime requirements suit the application.[5]