LFM3-Omni-2B-Uncensored

OpenIntelligenceNet/LFM3-Omni-2B-Uncensored is a surgically abliterated, uncensored version of LiquidAI/LFM3-Omni-2B created with Heretic (v2.0.0.dev0).

100% Preserved Omni-Modal Capabilities

To remove refusal behavior without degrading multimodal perception or speech synthesis, directional abliteration was performed exclusively on the 24-layer LFM3 language backbone (model.language_model) across 100 Optuna Pareto trials and transplanted back into the full Omni architecture.

• Untouched Vision Encoder (vision_tower + multi_modal_projector): Full image and video frame understanding.
• Untouched Audio Encoder (audio_encoder + audio_projector): Full 16kHz speech recognition (ASR) and video soundtrack understanding.
• Untouched Speech Synthesis (depth_decoder + LiquidCodec): Full 24kHz text-to-speech (TTS) and interleaved voice chat.
• Patched modeling_lfm3.py: Includes out-of-the-box support for linear_attention mask mappings across transformers releases.

---

Quickstart (Transformers)

import torch
from transformers import AutoModelForImageTextToText, AutoProcessor

model_id = "OpenIntelligenceNet/LFM3-Omni-2B-Uncensored"
device = "cuda" if torch.cuda.is_available() else "cpu"
dtype = torch.bfloat16 if device == "cuda" else torch.float32

processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForImageTextToText.from_pretrained(
    model_id,
    torch_dtype=dtype,
    device_map="auto",
    trust_remote_code=True
).eval()

# Text / Multimodal Chat Example
conversation = [
    {"role": "system", "content": "Respond with text."},
    {"role": "user", "content": "Explain how acoustic resonance works in musical instruments."}
]

inputs = processor.apply_chat_template(
    conversation,
    add_generation_prompt=True,
    tokenize=True,
    return_dict=True,
    return_tensors="pt"
).to(model.device)

with torch.no_grad():
    outputs = model.generate(**inputs, max_new_tokens=512, do_sample=False)

response = processor.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)
print(response)