Have you tried if it works with Qwen3-VL-2B ?
Hello, Just a question if you've tried with a qwen3-vl 2b encoder
Not tried, and right now it cannot be tried: ComfyUI has no Qwen3-VL-2B. QWEN3VL_CONFIGS in comfy/text_encoders/qwen3vl.py only declares 4b, 8b and 32b, so there is no architecture to load the checkpoint into.
Nothing in the method objects to it. The tokenizer is shared across the whole family, I checked that on 2014 texts and 472582 tokens, and the calibration script reads the student width from the checkpoint rather than assuming it. Someone would have to add the 2B config to ComfyUI first: hidden size, depth, deepstack indexes. After that the scripts run unchanged.
Expect it to be worse than the 4B, and not because of the projection. The 4B already misremembers things the 32B knows, and a 2B will misremember more. The matrix faithfully reproduces whatever the small model believes, mistakes included.
If you do add the config, run calibration/knows.py on it first. It asks the encoder to describe a few named people in plain text, which bypasses the matrix entirely and tells you where its knowledge stops. That takes two minutes and will tell you whether an hour of calibration is worth it.