Is there any plan to release a more lightweight model with fewer parameters?
Something similar to how Black Forest Labs released the Flux2 Klein model as a smaller version of Flux2.
That would be nice. However the text encoder would probably still remain the same large one, which is more of a problem depending on hardware. On my old system (4GB VRAM, 8GB RAM) the video model can run in about 20 minutes to generate a short clip, but before that the text encoder (GGUF Q2) runs for about 40 minutes on CPU which is unavoidable because it causes OOM on GPU.
That would be nice. However the text encoder would probably still remain the same large one, which is more of a problem depending on hardware. On my old system (4GB VRAM, 8GB RAM) the video model can run in about 20 minutes to generate a short clip, but before that the text encoder (GGUF Q2) runs for about 40 minutes on CPU which is unavoidable because it causes OOM on GPU.
I don’t think so. Flux2‑dev uses the text encoder mistral_3_small_flux2, while Flux2 Klein 9B uses qwen_3_8b.
Flux2‑dev uses the text encoder mistral_3_small_flux2, while Flux2 Klein 9B uses qwen_3_8b.
Ah, that's true. Then it's a matter of training a new, smaller model to work with another TE instead of just a reduction of model parameters within the same architecture. I was probably thinking of how earlier models such as Flex.1 and Flux.1 lite were pruned from the original model by the community.