Bagel fine-tune packs detection, OCR, depth and masks into one 7B-active model

mostlyired12 · reddit · 2026-09-27

SenseNova-Vision-7B-MoT is a fine-tune of ByteDance's Bagel, 14B total / 7B active, covering four vision tasks in one model:

Boxes and OCR return as plain text with coordinates; depth maps and masks come back as generated images.
On the paper's chosen comparisons it ranks first in object detection and OCR (one tie), beats Depth Anything V2 on every depth set, but trails MoGe-2 on most depth benchmarks, and dedicated segmentation models still win most mask tests.

Local deployment is hard: the only validated setup is a single 80GB A800, no GGUF exists, and llama.cpp can't load Bagel yet. Weights are non-commercial despite Bagel itself being Apache 2.0. The poster asks: if a GGUF got it onto 24GB, would you run one model for all four tasks or stick with Depth Anything + SAM + a small VLM?

Original post →

Bagel fine-tune packs detection, OCR, depth and masks into one 7B-active model

More from Models