base_model: FrostWolfAI/LFM3-1.7B-abliterated
base_model_relation: quantized
language:
- en
- ar
- zh
- fr
- de
- it
- ja
- ko
- pt
- es
- vi
- th
- id
- hi
- ru
- pl
license: other
license_name: lfm1.0
license_link: LICENSE
pipeline_tag: text-generation
tags: - gguf
- llama.cpp
- liquid
- lfm3
- edge
- on-device
- heretic
- ara
- uncensored
- decensored
- abliterated
- local-llm
LFM3-1.7B-abliterated-GGUF
GGUF quantizations of FrostWolfAI/LFM3-1.7B-abliterated
— an abliterated (decensored) LiquidAI/LFM3-1.7B produced with
Heretic v2 (ARA): 1/100 refusals at KL 0.0555 (base refuses 91/100).
Requires a recent llama.cpp build with
lfm3architecture support (also works in Ollama / LM Studio
builds that bundle a recent llama.cpp). Older builds will not load these files.
Quants
| File | Bits | Status |
|---|---|---|
LFM3-1.7B-abliterated-BF16.gguf |
16 | ✅ available (full precision) |
…-Q8_0 / Q6_K / Q5_K_M / Q4_K_M / Q4_K_S / Q3_K_M / Q2_K |
8–2 | 🔜 being added |
Smaller K-quants are being added. They require quantizing with a recent llama.cpp build that
supportslfm3(olderllama-quantizereportsunknown model architecture: 'lfm3'). Until then,
the BF16 GGUF runs on any recentlfm3-capable llama.cpp, or you can quantize it yourself.
Run
llama-cli -hf FrostWolfAI/LFM3-1.7B-abliterated-GGUF:Q4_K_M
# or local:
llama-server -m LFM3-1.7B-abliterated-Q4_K_M.gguf --jinja
LFM3 is a thinking model (<think>…</think>) with native tool-calling. Use --jinja so the chat
template (and tool-calling) render correctly.
License
LFM Open License v1.0 (lfm1.0), inherited from the base. See LICENSE. Refusal behavior
has been removed — released for security research and local use; you are responsible for your use.
Base: Liquid AI · Abliteration: Heretic by p-e-w · Build: FrostWolfAI Labs