Model

Qwen2.5-VL-3B

A multimodal vision-language model released by the Qwen team with 3 billion parameters.

Timeline (last 90 days)

Oct 20

SAPO achieves a 33.7% relative mIoU gain on Qwen2.5-VL-3B.

Unverified50%

Oct 12

Liquid has not published a comparison of d1-3B with Qwen2.5-VL-3B or Gemma, nor has it published calibration error data (such as ECE or reliability diagrams).

Unverified50%

All Facts (2)

Unverified

SAPO achieves a 33.7% relative mIoU gain on Qwen2.5-VL-3B.

50%Unverified

Liquid has not published a comparison of d1-3B with Qwen2.5-VL-3B or Gemma, nor has it published calibration error data (such as ECE or reliability diagrams).

50%

Source Articles

SAPO: Breaking the Video Temporal Grounding Bottleneck with Self-Supervised Attention Policy OptimizationOct 20