LongCat-Video-Avatar 1.5 — MLX INT8 fused

MLX-format weights for LongCat-Video-Avatar 1.5, converted for Apple Silicon with a fused INT8 DiT. This is the portable Natively Engine avatar pack.

Contents and installation

• base_model_int8_fused/: fused DiT and audio adapter. DMD is already merged; do not merge it a second time.
• vae/: full WAN VAE.
• lora/dmd_lora.safetensors: retained raw LoRA for source reproduction; ordinary fused inference does not download or use it.

Use natively models install longcat-int8 in the matching engine version. Its catalog separately installs the UMT5 tokenizer/text encoder and Whisper. No separate speech-generation model is needed when using an existing audio recording. Optional FlashVSR upscaling has its own catalog entry. LongCat INT4 is retired for output-quality degradation.

Provenance and terms

This notice-only update preserves all checkpoint bytes from revision 0883b3b38cec8821d0337f0cbcf8973b01d1b256; release-manifest.json records those public file identities. It does not retroactively claim an original source-to-conversion receipt for that legacy pack. The model family and merged DMD semantics are retained from the conversion and engine loader.

The Meituan LongCat Avatar 1.5 model card explicitly licenses model weights under MIT. Preserve the included LICENSE and Meituan LongCat attribution. MLX conversion and packaging: Shraey, Natively. Shared encoders keep their own notices. Natively Engine code is separately GPL-3.0-or-later.

Recorded engine benchmarks apply only to their source, runtime, settings and hardware. This metadata update changes no inference weights and makes no new speed, quality or minimum-memory claim.