FadedRedStar/LFM2.5-8B-A1B-heretic-GGUF

🤖 LFM2.5-8B-A1B-heretic — GGUF

This repository hosts GGUF weights for LFM2.5-8B-A1B-heretic, quantized from the source floating-point tensors provided by coder3101/LFM2.5-8B-A1B-heretic.

🔄 Sister Repository: Check out the Imatrix Sister Repository for enhanced precision at lower bit fractions.

If you plan on using 4-bit or 5-bit variants, consider the imatrix sister repository instead — importance matrix calibration improves logic retention at those bit depths. This repository is best suited if you want the near-lossless Q8_0 build.

ℹ️ Model Profile & Core Features

LFM2.5-8B-A1B is a text-only model from Liquid AI's Liquid Foundation Model 2.5 series, designed for on-device deployment. It uses a hybrid architecture with 24 layers — 18 double-gated LIV (Liquid, Input-adaptive, Value-selective) convolution layers plus 6 GQA (Grouped Query Attention) layers — activating only approximately 1.5B parameters per forward pass out of 8.3B total. This delivers fastest-in-class throughput at its size on both CPU and GPU, with day-one support for llama.cpp, MLX, vLLM, and SGLang. The model is a reasoning model: it produces a chain-of-thought before its final answer, and is tuned for complex instruction following, tool calling, and chained agentic task execution.

The heretic suffix denotes post-processing via the Heretic v1.2.0 Arbitrary-Rank Ablation (ARA) method with row-norm preservation performed by coder3101, which removes refusal conditioning at multiple tensor ranks while maintaining the model's instruction-following and planning capabilities.

📋 Technical Specifications

Property	Value
Base Architecture	LFM2.5 hybrid (18× double-gated LIV conv + 6× GQA)
Developed by	Liquid AI
Total Parameters	8.3B
Active Parameters	~1.5B per forward pass
Primary Use	Reasoning, instruction following, tool calling, agentic tasks
Context Window	128,000 tokens
Training Budget	38 trillion tokens
Languages	English, Arabic, Chinese, French, German, Japanese, Korean, Spanish, Portuguese
Abliteration Tool	Heretic v1.2.0
Abliteration Method	Arbitrary-Rank Ablation (ARA) with row-norm preservation
Prompt Format	ChatML

🛠️ Heretic Overrides (ARA)

Property	Value
start_layer_index	7
end_layer_index	21
preserve_good_behavior_weight	0.8548
steer_bad_behavior_weight	0.0004
overcorrect_relative_weight	0.9494
neighbor_count	8

📊 Refusal Bypass Metrics

The metrics below are self-reported by the original model author (coder3101) and have not been independently reproduced.

Metric	This model	Original (LiquidAI/LFM2.5-8B-A1B)
KL divergence	0.0239	0 (by definition)
Refusals	12/100	91/100

🧮 Numerical & Tensor Formats

Property	Value
Quantization Type	Q4_K_M, Q5_K_M, Q8_0

📦 Available Model Files

Main model weights

Filename	Quantization	llama.cpp Build	Size	Download
LFM2.5-8B-A1B-heretic-Q4_K_M.gguf	Q4_K_M	b9803	4.80 GB	📥 Download
LFM2.5-8B-A1B-heretic-Q5_K_M.gguf	Q5_K_M	b9870	5.62 GB	📥 Download
LFM2.5-8B-A1B-heretic-Q8_0.gguf	Q8_0	b9870	8.39 GB	📥 Download

🎛️ Component Pairing Guide

Download exactly one main weights file:

• Q4_K_M: Balanced 4-bit format suitable for most everyday use.
• Q5_K_M: Higher-fidelity mid-range format recommended as a general default.
• Q8_0: Near-lossless 8-bit format for when memory is not a constraint.

⚡ Deployment & Execution Commands

Liquid AI recommends the following generation parameters for best results: temperature: 0.2, top_k: 80, repetition_penalty: 1.05.

This model emits reasoning content before its final answer. If you require a clean final answer only, parse the output accordingly rather than expecting a single direct response.

Swap the -m filename below for either quantized file depending on your size/quality trade-off preference.

llama.cpp CLI

./llama-cli \
  -m LFM2.5-8B-A1B-heretic-Q4_K_M.gguf \
  -c 8192 \
  -ngl 99 \
  --temp 0.2 \
  --top-k 80 \
  --repeat-penalty 1.05 \
  -p "<|im_start|>system\nYou are a helpful and precise assistant capable of using tools and following complex instructions.<|im_end|>\n<|im_start|>user\nBreak down the following task and execute it step by step: summarise this document and list action items.<|im_end|>\n<|im_start|>assistant\n"

OpenAI-Compatible API Server

./llama-server \
  --host 0.0.0.0 \
  --port 8080 \
  -m LFM2.5-8B-A1B-heretic-Q4_K_M.gguf \
  -c 16384 \
  -ngl 99 \
  --flash-attn

💬 Chat Templates & Prompt Design (ChatML)

<|im_start|>system
You are a capable assistant. Follow instructions precisely.<|im_end|>
<|im_start|>user
Your task or query here.<|im_end|>
<|im_start|>assistant

⚠️ Safety & Operational Notes

• This model is abliterated and will generate content that standard aligned models refuse. Use responsibly and in compliance with applicable laws.
• This is a text-only model — it has no vision encoder and cannot process images.
• The LIV architecture activates only ~1.5B parameters per token, making it significantly faster to run than the total parameter count implies.
• For long-context workloads, set -c up to 131072 as needed.
• Liquid AI shipped a tokenizer fix for tool-calling after this model's initial release; if you encounter malformed tool-call output, verify your llama.cpp build includes this fix.
• For better output quality at this quantization level, consider the imatrix variant in the companion repository.