---
license: apache-2.0
base_model: DuoNeural/LFM2.5-8B-A1B-Abliterated
language:
- en
tags:
- gguf
- abliteration
- uncensored
- reasoning
- liquid-ai
- moe
- hybrid-attention
- quantized
- duoneural
pipeline_tag: text-generation
---

# LFM 2.5-8B-A1B Abliterated — GGUF Quantizations

GGUF quantizations of [DuoNeural/LFM2.5-8B-A1B-Abliterated](https://huggingface.co/DuoNeural/LFM2.5-8B-A1B-Abliterated), which is a DuoNeural abliteration of [Liquid AI's LFM 2.5-8B-A1B](https://huggingface.co/liquid-ai/lfm-2.5-8b-a1b).

**All capabilities fully preserved.** 100% factual accuracy, JSON generation, and math reasoning. 5/6 refusal behavior removed.

---

## Available Files

| Filename | Quantization | Size | RAM Required |
|----------|-------------|------|--------------|
| `lfm25-8b-abliterated-q5_k_m.gguf` | Q5_K_M | 5.7 GB | ~8 GB |
| `lfm25-8b-abliterated-q4_k_m.gguf` | Q4_K_M | 4.9 GB | ~6 GB |

**Recommended:** Q5_K_M for best quality/size balance.

## Inference Speed Benchmarks

| Hardware | Quantization | Generation | Prefill | Context |
|----------|-------------|------------|---------|---------|
| NVIDIA GTX 1070 (8GB) | Q4_K_M | **63.88 t/s** | 0.41s | 128k (full) ✓ |
| NVIDIA GTX 1070 (8GB) | Q5_K_M | **31.05 t/s** | 0.70s | 64k confirmed ✓ |
| NVIDIA RTX 3090 (pod) | Q4_K_M | **388 t/s** | ~250 t/s | 128k ✓ |

The GTX 1070 results are remarkable — a **10-year-old** consumer GPU (released June 2016) fully loading the 128k context window of a reasoning model at 63 t/s. This model is efficient.

---

## Usage with llama.cpp

```bash
# Download
huggingface-cli download DuoNeural/LFM2.5-8B-A1B-Abliterated-GGUF \
  lfm25-8b-abliterated-q5_k_m.gguf

# Run
./llama-cli -m lfm25-8b-abliterated-q5_k_m.gguf \
  --chat-template chatml \
  -p "Your prompt here" \
  -n 512
```

## Usage with Ollama / LM Studio

Import the GGUF file directly. The model uses ChatML template format.

---

## KL Divergence (Distribution Shift)

Measured via Heretic v1.2.0 methodology (100 benign prompts, full vocab, first-token logits):

| Metric | Value |
|--------|-------|
| KL(abliterated \|\| original) | **1.06 × 10⁻⁷ nats** |

Near-zero KL on benign prompts confirms the orthogonal projection is surgical: the refusal direction is minimally activated on benign inputs, leaving factual/coding/math generation untouched. See the [BF16 model card](https://huggingface.co/DuoNeural/LFM2.5-8B-A1B-Abliterated) for full methodology comparison vs LoRA-based abliteration.

---

## About the Abliteration

See [DuoNeural/LFM2.5-8B-A1B-Abliterated](https://huggingface.co/DuoNeural/LFM2.5-8B-A1B-Abliterated) for full methodology and validation results.

**Quick summary:**
- Targets all 6 full-attention GQA layers (positions 2, 6, 10, 14, 18, 21)
- alpha=1.5 for reasoning model (higher than standard due to `<think>` CoT architecture)
- Correct WRITER and READER projection formulas
- 5/6 test prompts answered post-abliteration; original answered ~1/6
- Zero capability degradation

> **Note:** Model card will be updated with paper citation when the associated DuoNeural research paper on Dual Horizon Processing in Hybrid Architectures is published.

---

## About DuoNeural

**DuoNeural** is an open AI research lab operating at the intersection of human and artificial intelligence. We study post-training dynamics, mechanistic interpretability, temporal sequence learning, and quantum machine learning — publishing everything under open access.

📄 **Full paper catalog:** [zenodo.org/communities/duoneural](https://zenodo.org/communities/duoneural)  
🤗 **HuggingFace:** [huggingface.co/DuoNeural](https://huggingface.co/DuoNeural)  
🌐 **Website:** [duoneural.com](https://duoneural.com)  

*All research published open access, CC BY 4.0.*