The images behind the story
ByteDance’s Cola2 beats Llama 3 on two tests. What do the other six show?
The Cola2 research checkpoint edges past Llama 3 8B on two continuation tasks. Its full results and unusual decoding method explain what developers can reproduce.
Article by GPT · Read the full story →
The editorial illustration
AI-generated editorial illustration of semantic compression and text decoding.
- Image model
- google/gemini-3.1-flash-lite-image
- Why this image
- The folded sections and individual tiles illustrate two levels of representation without implying they are an actual model visualization.
Read the full image prompt
Editorial photograph-style paper sculpture, wide landscape 3:2. On a dark teal worktable a flowing ribbon of tiny cream paper tiles gathers into several distinctly colored folded bundles, then unfolds into a new ribbon of individual tiles. Strong simple composition, warm sidelight, tactile paper grain. Visual metaphor for compressing phrases and decoding words. No type, letters, numbers, logos, charts or humans.
Produced by Maker Frontier · Image use & permissions
The image-generation model credited above made this illustration. It may differ from the model discussed in the article.
Code-rendered visual summary; AI-designed layout. Text comes from the article headline and summary.
View this image and its source →Produced by Maker Frontier · Image use & permissions