klein 9B pairs with Qwen3-8B while klein 4B pairs with Qwen3-4B, and handing a klein 9B
checkpoint the 4B encoder fails deep inside sd-cli, so Unsloth Studio picks this file per variant
on the no-GPU path. It used to come straight from the community repack, which carries no model
card and no licence metadata at all; a rename or a takedown there turns every CPU klein-9B load
into a 404, so it is mirrored here. The weights are unmodified: byte for byte the file of the
same name in the source repo.
Only this file is mirrored. The source repo also ships a FLUX.2 VAE and quantised conditioners;
Studio does not use them from there, so they are not here.
Licence: Apache-2.0, and here is the evidence
This file ships inside black-forest-labs/FLUX.2-klein-9B, which is under the FLUX
Non-Commercial License, so the obvious move would be to label it non-commercial too. That would
be wrong: the weights are stock Qwen3-8B, with nothing of Black Forest Labs's in them.
All 399 tensors were compared against Qwen/Qwen3-8B by byte hash of the raw tensor buffer, key
for key:
BIT-IDENTICAL 399 DIFFERENT 0
Same key names, same dtypes, same shapes, no key in one file missing from the other. The FLUX
Non-Commercial License itself provides at section 9 that third-party components stay subject to
their own licensors' terms, so Alibaba's Apache-2.0 grant is what governs these bytes. Full text
in LICENSE,
attribution and the statement of changes in
NOTICE.
This says nothing about FLUX.2 klein 9B itself. That model is under the FLUX Non-Commercial
License, and running it is governed by those terms no matter what licence its conditioner carries.
Not an official Alibaba Cloud, Black Forest Labs or ComfyUI product, and not endorsed by any of
them.
