Qwen-VL generational lineage — 2023 to 2025
2023
Sep 2024
Feb 2025
Nov 2025
Qwen-VL
OpenCLIP + Q-Former
448x448 resolution
grounding + bbox tokens
zh + en multilingual
Qwen2-VL
M-RoPE (3D position)
native dynamic res
MLP projector only
video with variable FPS
Qwen2.5-VL
dynamic FPS sampler
absolute-time M-RoPE
window attention ViT
JSON agent mode
Qwen3-VL
Qwen3-72B base
thinking mode
expanded OCR
data + scale
M-RoPE: one position, three axes
temporal band
16 dims (out of 48)
t=0 image, t=frame video
rotate by t * theta_i
height band
16 dims
h = patch row index
rotate by h * theta_i
width band
16 dims
w = patch column index
rotate by w * theta_i
dynamic FPS: tokens per video = duration * fps * tokens_per_frame
high motion
tennis, cooking
4-8 FPS
event-dense
~19k tokens / 30s
medium motion
dialogue, demo
2-4 FPS
balanced
~10k tokens / 30s
low motion
security, lecture
0.5-1 FPS
long-horizon
~24k tokens / 10min
agent replay
UI screencasts
2-4 FPS
JSON output mode
click/scroll calls