Iris-3B Turbo MLX Q4
Complete, unofficial community r2-1000 Turbo LoRA-merged package. This is distinct from Sperid's official full-weight Turbo release. The validated image checkpoint retains pixel/input/modulation layers in BF16 and quantizes compatible trunk projections to affine Q4/group64. The language-only Qwen3-VL-4B text encoder is also Q4/group64; unused visual tensors are omitted. Model licenses and source documentation are retained under sources/.
Native Swift/MLX: zimageapp/zimage, IrisCLI --model /path/to/package --prompt "a red apple" --steps 4 --shift 4 --seed 42 --output out.png.
The sampler uses 4 stochastic Turbo steps and predicts pixels directly, without a VAE. --fp16-dense selects trunk projections; --cache-dit enables approximate prediction reuse, off by default. Source image/text precisions and quantization parameters are explicit in the package. Text Q4 changes conditioning relative to BF16 text; consult the repository's native model report for the measured comparison.
Quantized