Ant Group's UniSpark-9B unifies multimodal understanding and generation with hybrid attention

antgroup · hf · 2026-11-13

Ant Group released a technical report on UniSpark-9B, a unified multimodal model for visual understanding and generation:

Architecture: Mixture-of-Transformer-Experts backbone with 9B parameters per branch, combining full attention for explicit token interactions with GatedDeltaNet linear attention for compact context propagation.
Capabilities: autoregressive text prediction plus visual flow matching support VQA, text-to-image generation, instruction-guided editing, and interleaved text-image prediction; a blockwise protocol isolates noisy visual targets during interleaved generation via attention masks and recurrent-state retention rules.
Training: 159.7M generation samples, 213.4M understanding samples, 624.8K interleaved sequences, with a four-stage curriculum extending image synthesis to 2K resolution.
Results: 2470.3 on MME-S and 0.83 on GenEval, plus qualitative demos of interleaved image generation, video synthesis, and video editing.

More from Multimodal