MLXBits/ltx-2.3-10eros-v1.3-dmd-mlx-q4

LTX-2.3 10Eros v1.3 DMD — MLX int4 (q4)

Int4 (4-bit) MLX quantization of **TenStrip/LTX2.3-10Eros** v1.3 with the JoyAI Echo DMD distillation baked directly into the transformer, replacing the native rank-384 distilled LoRA.

Converted with mlx-forge and packaged for **ltx-2-mlx** — a pure-MLX port of LTX-2 for Apple Silicon.

[!WARNING] Not for all audiences. 10Eros is intended for adult use. By downloading and running this model you confirm you are of legal age in your jurisdiction and accept responsibility for the content you generate. Do not use it to produce illegal material or to depict real, identifiable people without consent.

[!NOTE] This is the int4 build — the smallest and lightest variant, roughly half the transformer footprint of the int8 (q8) sibling. Int4 weight quantization trades some fine detail and prompt-adherence fidelity for the smaller size. If quality matters more than footprint, prefer the int8 build.

What this is

This is a distilled-only package. The DMD distillation deltas extracted from JoyAI Echo (rank-256, reshaped from the LTX 384 1.1 distilled-LoRA shapes, audio branch sourced from the 384 1.1 distilled LoRA) were merged into the 10Eros v1.3 base at strength 1.0, producing a standalone distilled transformer. There is no dev transformer and no runtime LoRA fusion — the distillation is already in the weights.

Compared to the native 384 distilled LoRA, the DMD merge avoids the resampling-to-base drift, the conditioning drop from the latent, and the extra detail/look overwrites that the stock distilled LoRA introduces at the upscale refine stage.

Run it with `--distilled`. The two-stage (--two-stage, --two-stages-hq) and one-stage-dev (--one-stage) paths are not available here — they require a dev transformer, which this package intentionally omits.

What's in here

Quantization: int4, group size 32, applied only to nn.Linear inside transformer_blocks. AdaLN, projections, connectors, VAE and vocoder remain bf16 (MLX cannot quantize Conv layers). At int4 the transformer is ~11.6 GB versus ~19 GB at int8 — the lightest way to run this model.

Text encoder: Gemma 3 12B is not bundled — ltx-2-mlx loads it separately via mlx-lm.

[!NOTE] mlx-forge may also drop ltx-2.3-22b-distilled-lora-384*.safetensors (~7.6 GB each) into this directory as "shared" LoRA components. They are unused in a distilled-only package — nothing fuses them (there is no dev transformer, and --distilled never loads a LoRA). Safe to delete; also remove them from the lora list in split_model.json.

Usage

Requires ltx-2-mlx on Apple Silicon.

# Text-to-video (distilled two-stage: half-res → upscale → full-res refine)
ltx-2-mlx generate \
  --model /path/to/ltx-2.3-10eros-v1.3-dmd-mlx-q4 \
  --prompt "your prompt" \
  --distilled \
  -H 480 -W 704 -f 97 -o out.mp4

# Image-to-video — add --image
ltx-2-mlx generate \
  --model /path/to/ltx-2.3-10eros-v1.3-dmd-mlx-q4 \
  --prompt "animate this" \
  --distilled --image photo.jpg -o out.mp4

Default distilled flow is the usual 8/4-step upscale schedule. Experiment with other sigmas or any Euler / LTX-compatible sampler — no custom loading or sampling is needed. The int4 transformer roughly halves the resident weight footprint versus int8, so this is the most memory-friendly build; on 16 GB Macs add --low-ram for block-streamed inference.

Conversion provenance

1. 1.The JoyAI Echo DMD LoRA was merged into the 10Eros v1.3 bf16 base at per-layer strength 1.0 (alpha ÷ actual rank) to produce a distilled bf16 checkpoint:

   uv run ~/mlx-forge/scripts/merge_lora.py
   # LoRA:   LTX2.3_DMD_reshaped_r256.safetensors
   # → /Volumes/Storage/10Eros-v1.3-distilled-dmd-bf16.safetensors

1. 1.Distilled variant + all shared components + upscalers, quantized to int4 (group size 32):

   uv run mlx-forge convert ltx-2.3 --variant distilled \
     --checkpoint /Volumes/Storage/10Eros-v1.3-distilled-dmd-bf16.safetensors \
     --quantize --bits 4 --group-size 32 \
     --spatial-upscaler x2 x1.5 --temporal-upscaler x2 \
     --output models/ltx-2.3-10eros-v1.3-dmd-mlx-q4

No dev variant is added — this package is distilled-only by design.

Credits & license

• —DMD distillation: JoyAI Echo — 10S-Comfy-nodes
• —Finetune: TenStrip/LTX2.3-10Eros
• —Base model: Lightricks/LTX-2.3
• —Conversion: mlx-forge · Runtime: ltx-2-mlx

Distributed under the LTX-2 license. The 10Eros finetune's own terms also apply — review the upstream model card before redistribution.