NOTE: weights are now bf16, not fp8. Ideogram-4 with the twinks_10000 LoKr merged in at full bf16 precision (crisp, no fp8 requantization). Layout matches ideogram-ai/ideogram-4-fp8 (transformer/ + unconditional_transformer/), but each diffusion_pytorch_model.safetensors is bf16, single-shard, 458 tensors, no weight_scale. Text encoder / VAE / tokenizer / scheduler are the stock Ideogram-4 files. For AI Toolkit, set the Ideogram model name_or_path to this repo.