FLUX.2 [klein] 4B, float8 weights

This is a modified copy of black-forest-labs/FLUX.2-klein-4B by Black Forest Labs, released under the Apache License 2.0 (see LICENSE.md).

What changed: the large linear layers of transformer/ and text_encoder/ were cast from bfloat16 to float8 (e4m3fn) without scaling, and re-sharded. Layers matching norm, embed, proj_out, x_embedder, context_embedder, time, lm_head stay in bfloat16, and so do the VAE, tokenizer, scheduler and configs. No weights were trained or tuned. mcr_fp8.json records the exact settings.

Load the layers with float8 storage and bfloat16 compute (for example diffusers' apply_layerwise_casting), as Minecraft Clip Radar's Thumbnail Lab does.

Downloads last month
20
Safetensors
Model size
4B params
Tensor type
BF16
ยท
F8_E4M3
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for AwesomeEZ12/FLUX.2-klein-4B-fp8

Finetuned
(76)
this model