Agate brings SDXL-level GenEval to a 260M-parameter model
Agate Preview 001 gives developers an MIT-licensed text-to-image model that matches SDXL's published 0.55 GenEval score in 260M parameters, making local experimentation far lighter. Its small transformer “thinker” turns a prompt into a 16 × 16 region plan, then steers a convolutional renderer across the image. The model generates a 256 × 256 image in 1.9 seconds on an RTX 4060, but remains a research preview with weak exact text and counting.