Sign up

xocialize/qwen-image-edit-2511-mlx-int4

🤗 Hugging Face sourceimage-to-imageapache-2.019 GBother✓ Checksums witnessedupdated today
No torrent yet →

qwen-image-edit-2511-mlx-int4

Pre-quantized MLX weights for Qwen-Image-Edit-2511, for the Swift xocialize/qwen-image-edit-swift package's QwenImageEdit turbo tier. They let the engine load the int4 tier directly instead of loading bf16 (~40 GB peak) and quantizing in-memory.

⚠️ Non-standard format — for the MLXEngine Swift loaders only. These are not standard mlx_lm / mlx_vlm / diffusers-layout repos and will not load with generic MLX tooling. Tensors use the engine's internal module key names and carry custom quantization metadata; they load only via QwenImageEditWeights.loadQuantizedDiT and QwenVLPromptEncoder.load(quantizedTextModelPath:). For a standard, tool-loadable conversion see mlx-community/qwen-image-edit-2511-8bit.

Files

file size contents
qie-2511-dit-int4-mod8.safetensors ~15 GB 20B DiT transformer — attention + feed-forward Linears int4 (group 64), modulation linears int8 (conditioning-critical), top-level projections bf16. Per-layer bits + group size travel in the safetensors metadata.
qie-2511-vl7b-int4.safetensors ~4 GB Qwen2.5-VL-7B text model, int4 (group 64). The (small) vision tower is not included — it loads bf16 from the base snapshot.

Measured footprint (Swift MLXEngine turbo tier, 1024², 4-step DMD)

  • Load peak ~21 GB (no bf16 ever materialized) · resident ~21 GB · inference peak ~25.5 GB (with bf16 VAE).
  • vs ~57 GB bf16 / ~41 GB quantize-after-load peak.

Provenance / license

Derived from Qwen/Qwen-Image-Edit-2511 (Apache-2.0). Quantization: MLX affine, group size 64. The Lightning 4-step DMD LoRA is applied separately at runtime and is not included here.