LFM3-1.7B Heretic Uncensored — GGUF
A decensored version of Liquid AI's LFM3-1.7B, converted to GGUF for efficient local inference.
This repository provides multiple GGUF precision variants of Peymandc/LFM3-1.7B-Heretic-Uncensored, making the model accessible through compatible inference engines such as llama.cpp and other applications supporting LFM3 GGUF models.
All GGUF files in this repository were converted from the modified Safetensors checkpoint, preserving the Heretic modifications. They are not conversions of the original, unmodified Liquid AI checkpoint.
The original modified Safetensors checkpoint is available here:
Peymandc/LFM3-1.7B-Heretic-Uncensored
Available GGUF Versions
| Quantization | File Size | Description |
|---|---|---|
| Q4_K_M | 1.10 GB | Recommended balance between model quality and memory usage. |
| Q6_K | 1.43 GB | Higher precision with greater memory requirements. |
| Q8_0 | 1.85 GB | High-quality 8-bit quantization with minimal quantization loss. |
| BF16 | 3.47 GB | BF16 weight representation, retaining the source checkpoint's original weight precision. |
Download
- LFM3-1.7B Heretic.Q4_K_M.gguf
- LFM3-1.7B Heretic.Q6_K.gguf
- LFM3-1.7B Heretic.Q8_0.gguf
- LFM3-1.7B Heretic.BF16.gguf
Choose the quantization that best fits your available RAM, VRAM, and performance requirements.
For most users looking for a balance between quality and resource consumption, Q4_K_M is a good starting point. Q6_K and Q8_0 offer higher weight precision, while BF16 is intended for systems with sufficient memory.
File sizes are approximate and reflect the sizes displayed by Hugging Face.
No comparative benchmark results are currently claimed for these GGUF variants.
About the Model
This release is based on LiquidAI/LFM3-1.7B, modified using Heretic v2.1.0 to reduce the original model's refusal behavior.
The modification was performed using Heretic's Arbitrary-Rank Ablation (ARA) approach. The objective is to retain the model's underlying capabilities while changing its learned refusal behavior.
This is a refusal-behavior modification, not a separately trained foundation model.
💜 Enjoying my work? You can support my independent AI projects here.
Heretic Modification
The original Safetensors release documents the following Heretic configuration:
| Parameter | Value |
|---|---|
| Heretic version | 2.1.0 |
| Selected optimization trial | #97 |
| Start layer | 9 |
| End layer | 24 |
| Preserve good behavior weight | 0.3616 |
| Steer bad behavior weight | 0.0006 |
| Overcorrect relative weight | 0.7002 |
| Neighbor count | 14 |
Heretic Evaluation
The source checkpoint's Heretic evaluation reported:
| Metric | Modified Model | Original LFM3-1.7B |
|---|---|---|
| Refusal-keyword matches | 3/100 (3%) | 98/100 (98%) |
| KL divergence | 0.0732 | 0.0000 |
Evaluation methodology
- Refusal-keyword matches: Heretic's
KeywordRatescorer, evaluated on the first 100 prompts frommlabonne/harmful_behaviors(test[:100]). A response is counted when it contains any configured refusal keyword, case-insensitively. This is a keyword-based proxy, not a semantic or human-reviewed refusal evaluation. - KL divergence: Heretic's
KLDivergencescorer, evaluated on the first 100 prompts frommlabonne/harmless_alpaca(test[:100]). The original model is the reference, so its divergence is zero by definition.
Full evaluation settings, keyword markers, dependencies, optimization history, and reproduction instructions are available in the original Safetensors repository's reproduce/ directory.
These results apply to the evaluated source checkpoint. The individual GGUF quantizations have not been independently evaluated.
Model Specifications
| Property | Specification |
|---|---|
| Model | LFM3-1.7B Heretic Uncensored |
| Architecture | Tri-hybrid |
| Parameters | Approximately 1.73B |
| Model type | Post-trained, text-only language model |
| Layers | 24 |
| Vocabulary | 128,000 tokens |
| Context length | Up to 256K tokens (input + output) |
| Original developer | Liquid AI |
| Base checkpoint | LiquidAI/LFM3-1.7B |
| Modification | Heretic v2.1.0 |
| Format | GGUF |
| Weight formats | Q4_K_M, Q6_K, Q8_0, BF16 |
| License | LFM Open License v1.0 |
Supported Languages
The original LFM3-1.7B documentation lists support for the following 16 languages:
- English
- Arabic
- Chinese
- French
- German
- Italian
- Japanese
- Korean
- Portuguese
- Spanish
- Vietnamese
- Thai
- Indonesian
- Hindi
- Russian
- Polish
This language list is inherited from the original LFM3-1.7B documentation. Multilingual capabilities have not been independently evaluated on this Heretic variant.
Architecture Overview
The underlying LFM3 architecture combines three types of processing:
- 13 gated short-convolution layers
- 5 sliding-window attention layers
- 6 global attention layers
This tri-hybrid architecture combines efficient local processing with attention-based mechanisms for handling broader contextual information.
The 256K context length is inherited from the original model architecture. Actual usable context depends on the inference engine, hardware, memory availability, and runtime configuration.
Intended Use
This release is intended for:
- Local and offline AI inference.
- Conversational assistants and general text generation.
- Experimentation with refusal-modified language models.
- Personal AI applications and edge devices.
- Research and evaluation of lightweight language models.
Although the original LFM3 model supports agentic workflows, tool calling, and structured outputs, the capabilities of this modified release should be independently evaluated for each application.
As a relatively small 1.73B-parameter model, it should not be expected to match larger models in complex reasoning, extensive knowledge tasks, or demanding multi-step problem solving. It is not recommended for demanding agentic coding or complex multi-step reasoning.
Usage
These GGUF files are intended for inference engines with LFM3 architecture support.
Recommended inference engine: llama.cpp
Make sure you use a recent llama.cpp build that supports LFM3. Older builds may not recognize the architecture or correctly handle its chat template.
Example Using llama.cpp
llama-cli -m "path/to/model.gguf" -cnv
The GGUF file should be loaded with its supported chat template. Use the inference engine's built-in template handling whenever available.
For more information about the original model's chat template, reasoning format, and tool-calling capabilities, refer to the main Safetensors repository.
Limitations and Responsible Use
This model has undergone refusal-behavior modification using Heretic. As a result, it may respond to requests that the original model would have refused.
It may generate inaccurate, misleading, harmful, or otherwise inappropriate content. Its outputs should not be considered verified information or professional advice.
Users are responsible for evaluating generated content and ensuring that their applications comply with applicable laws and regulations.
License
This model is distributed under the LFM Open License v1.0, inherited from the original Liquid AI model.
Please review the official license before using or redistributing the model.
Commercial use is subject to the restrictions specified in Section 5 of the license.
Acknowledgements
This release is derived from LiquidAI/LFM3-1.7B, developed by Liquid AI.
Refusal-behavior modification was performed using Heretic.
The GGUF variants are provided by Peymandc.
Please credit the original model developers and the modified release when using this work.
Explore more models by Peymandc
💜 Support Independent Development
These projects are part of my ongoing exploration of open-weight AI models.
Your support helps make more room for experimentation, model modifications, testing, research, and future releases.
If you find my work useful and would like to contribute, you're welcome to visit my support page:
Every contribution is appreciated, but never expected. Sharing feedback, reporting issues, and simply using the models are also wonderful ways to support independent development.
Thank you for being here. ❤️