10eros-experimental-mlx-q8

10Eros Experimental MLX Q8

10Eros v1 for Apple Silicon via ltx-2-mlx, with the Sulphur experimental LoRA fused at 0.35 strength.

This is an 8-bit quantization, targeted at 32GB+ Macs running Apple Silicon. If there is any interest in a 4-bit quant, use the community tab to request one, noting your Mac's spec. In my experience, 32GB M-series is the low bar for using anything based on LTX2.3.

Model Contents

How It Was Built

Step 1 — Sulphur merge (bf16)

The Sulphur experimental LoRA was merged into 10Eros_v1_bf16.safetensors at strength 0.35 using a custom merge_lora.py script. This produces the dev transformer source.

Step 2 — Condsafe distilled merge (bf16)

TenStrip's condsafe distilled LoRA (ltx-2.3-22b-distilled-lora-1.1_fro90_ceil72_condsafe.safetensors) was merged into the Step 1 output at strength 1.0. This produces the distilled transformer source.

The standard distilled-384 LoRAs from Lightricks are incompatible with 10Eros — they produce a frosted-glass texture artifact. TenStrip's condsafe LoRA is specifically designed to work with this finetune.

Step 3 — MLX conversion + quantization

Converted with mlx-forge:

# Distilled transformer + shared components
mlx-forge convert ltx-2.3 \
  --checkpoint 10eros-experimental-distilled-bf16.safetensors \
  --variant distilled-1.1 --quantize --bits 8 \
  --spatial-upscaler x2 x1.5 --temporal-upscaler x2 \
  --output 10eros-experimental-mlx-q8

# Dev transformer (added to same directory)
mlx-forge convert ltx-2.3 \
  --checkpoint 10eros-experimental-bf16.safetensors \
  --variant dev --quantize --bits 8 --skip-shared \
  --output 10eros-experimental-mlx-q8

Usage

# Distilled (fast, recommended)
ltx-2-mlx generate --model MLXBits/10eros-experimental-mlx-q8 \
  --distilled --prompt "your prompt" --low-ram -o output.mp4

# One-stage dev (slower, CFG)
ltx-2-mlx generate --model MLXBits/10eros-experimental-mlx-q8 \
  --one-stage --prompt "your prompt" --low-ram -o output.mp4

Credits

• 10Eros v1: TenStrip/LTX2.3-10Eros
• Sulphur 2 Experimental LoRA: SulphurAI/Sulphur-2-base
• MLX conversion: mlx-forge
• Inference: ltx-2-mlx