FLUX.2 starter bundle installs gated, non-commercial Klein 9B and FLUX.2 [dev]

[lstein]

Summary

Since #9656 (commit 91db6b8), the FLUX.2 starter bundle installs FLUX.2 Klein 9B and FLUX.2 [dev] alongside Klein 4B. Both added models are under Black Forest Labs' non-commercial licenses, and the Klein 9B text encoder downloads from a gated repository. The bundle used to be the Apache 2.0, ungated Klein 4B set, so one click on the Launchpad or in Add Models now:

• installs non-commercial weights, with nothing in the UI saying so;
• fails partway for anyone without a Hugging Face token, or who hasn't accepted the Klein 9B license;
• downloads about 70 GB instead of about 11 GB.

The code comment above the bundle is also wrong about the license.

Details

invokeai/backend/model_manager/starter_models/__init__.py (on main at 587fabb):

# Klein 4B and 9B (each with its Qwen3 encoder), the PiD super-resolution decoder for their latents, and FLUX.2 [dev]
# at GGUF Q4 with its cow-mistral encoder. [dev] is non-commercial; the Klein models are not.
flux2_bundle: list[StarterModel] = [
    flux2_klein_4b_gguf_q4,
    flux2_klein_qwen3_4b_encoder,
    flux2_klein_9b_gguf_q4,
    flux2_klein_qwen3_8b_encoder,
    flux2_vae,
    pid_decoder_flux2_2k,
    flux2_dev_gguf_q4_k_m,
    flux2_dev_cow_mistral_q4,
]

The comment's claim that "the Klein models are not" non-commercial contradicts our own docs. The license table in docs/src/content/docs/Users Guide/03.Models/Local Models/flux-2.mdx says:

Model	License	Gated
FLUX.2 Klein 4B	Apache 2.0	No
FLUX.2 Klein 9B	FLUX Non-Commercial License	Yes, on Black Forest Labs' repositories
FLUX.2 [dev]	FLUX [dev] Non-Commercial License	Yes, on Black Forest Labs' repositories

What the bundle now pulls in:

Starter model	Source	Size
Klein 4B GGUF Q4	unsloth/FLUX.2-klein-4B-GGUF	~2.6 GB
Qwen3 4B encoder	black-forest-labs/FLUX.2-klein-4B	~8 GB
Klein 9B GGUF Q4 (non-commercial)	unsloth/FLUX.2-klein-9B-GGUF	~5.8 GB
Qwen3 8B encoder	black-forest-labs/FLUX.2-klein-9B (gated)	~16 GB
FLUX.2 VAE	black-forest-labs/FLUX.2-klein-4B	~0.2 GB
PiD 2K decoder	nvidia/PiD	~5 GB
FLUX.2 [dev] Q4_K_M (non-commercial)	gguf-org/flux2-dev-gguf	~20 GB
cow-mistral3-small Q4	gguf-org/flux2-dev-gguf	~11.6 GB

The sizes above add up to about 69 GB, before PiD's own dependencies are counted.

Steps to reproduce

1. Without a Hugging Face token, or without having accepted the Klein 9B license, open the Launchpad or Add Models and install the FLUX.2 bundle.
2. The Qwen3 8B encoder install fails on the gated black-forest-labs/FLUX.2-klein-9B repo. Klein 9B (~5.8 GB) still downloads from unsloth but cannot run without that encoder.
3. With a token, the bundle installs FLUX.2 [dev] and Klein 9B, which are non-commercial, as part of what looks like the default FLUX.2 setup. Nothing in the UI mentions the license.

Expected

A default bundle for a base model should contain only models that are ungated and can be used commercially, or at least clearly say when it doesn't.

Suggested fix

• Restore the FLUX.2 bundle to the ungated, Apache 2.0 Klein 4B set: Klein 4B, the Qwen3 4B encoder, the FLUX.2 VAE, and optionally the PiD decoder.
• Keep Klein 9B and FLUX.2 [dev] installable on their own, or put them in a separate bundle whose name or description says it is gated and non-commercial.
• Correct the code comment to match the license table in flux-2.mdx.
• Optionally, add a test asserting that bundle members do not come from known gated or non-commercial sources, so this can't come back silently.

Related: the user docs still describe the pre-#9656 bundle names and contents. flux-2.mdx still calls the bundle "FLUX.2 Klein"; the Wan, LTX, FLUX.1, Z-Image and SD 1.5 pages have the same problem.