language:

• en
  library_name: mlx
  base_model: OpenIntelligenceNet/LFM3-Omni-2B-Uncensored
  pipeline_tag: image-text-to-text
  tags:
• mlx
• mlx-vlm
• osaurus
• vmlx
• lfm3
• lfm3-omni
• multimodal
• vision
• audio
• 8-bit
• abliterated
• uncensored

LFM3-Omni-2B-Uncensored-MLX-8bit

Near-lossless 8-bit MLX conversion of OpenIntelligenceNet/LFM3-Omni-2B-Uncensored for Apple Silicon, built with mlx-vlm (0.7.11) and verified for out-of-the-box compatibility with Osaurus (vmlx) and mlx-vlm.

• 4-bit sibling repo: spiderversekid/LFM3-Omni-2B-Uncensored-MLX-4bit

Conversion & Quantization Details

• Language Backbone (language_model): Preserves the original abliterated weights from OpenIntelligenceNet/LFM3-Omni-2B-Uncensored without re-abliteration:
  • Decoder linear, short-conv, and embedding projections: 8-bit affine (group_size=64) (0.99995 logit cosine similarity and 100% top-1 token agreement vs. bfloat16)
  • Attention head gates (self_attn.gate_proj, [16, 2048]): Full bfloat16 precision
• Multimodal & Audio Modules (Full bfloat16 Precision):
  • Vision Tower (vision_tower + multi_modal_projector)
  • Audio Encoder (audio_tower + audio_projector)
  • Audio Generation (depth_decoder + codebook_embedding)
  • Liquid Audio Codec (audio_tokenizer/model.safetensors)

Usage

1. Osaurus (vmlx)

Import spiderversekid/LFM3-Omni-2B-Uncensored-MLX-8bit directly in Osaurus or place it in ~/MLXModels/spiderversekid/LFM3-Omni-2B-Uncensored-MLX-8bit.

2. mlx-vlm CLI

pip install -U mlx-vlm

Text-only generation:

python -m mlx_vlm.generate \
  --model spiderversekid/LFM3-Omni-2B-Uncensored-MLX-8bit \
  --prompt "Explain what a multimodal omni model can do." \
  --max-tokens 256 \
  --temperature 0.0

Image + Text generation:

python -m mlx_vlm.generate \
  --model spiderversekid/LFM3-Omni-2B-Uncensored-MLX-8bit \
  --image path/to/image.png \
  --prompt "Describe this image in detail." \
  --max-tokens 256 \
  --temperature 0.0