Requires a recent llama.cpp build with LFM2 MoE support. This model uses the lfm2moe architecture (hybrid short-conv + attention with 32 experts, 4 active per token). Only llama.cpp builds that include Lfm2MoeForCausalLM support can load these files. Use a current release (or current Ollama / LM Studio). Older builds will fail with an "unknown architecture 'lfm2moe'" error.

Files

File	Quant	Size	BPW	Notes
`LFM2.5-8B-A1B-Uncensored-Q2_K.gguf`	Q2_K	3.0 GB	3.01	Smallest; significant quality loss but works on very constrained hardware.
`LFM2.5-8B-A1B-Uncensored-Q3_K_S.gguf`	Q3_K_S	3.5 GB	3.54	Small, lower quality.
`LFM2.5-8B-A1B-Uncensored-Q3_K_M.gguf`	Q3_K_M	3.9 GB	3.87	Small; some quality loss.
`LFM2.5-8B-A1B-Uncensored-IQ4_XS.gguf`	IQ4_XS	4.3 GB	4.25	Smaller than Q4_K_S with comparable quality; uses iquant scheme.
`LFM2.5-8B-A1B-Uncensored-Q4_K_S.gguf`	Q4_K_S	4.6 GB	4.59	Slightly smaller than Q4_K_M.
`LFM2.5-8B-A1B-Uncensored-Q4_K_M.gguf`	Q4_K_M	4.9 GB	4.85	Recommended — best size/quality balance for most users.
`LFM2.5-8B-A1B-Uncensored-Q5_K_S.gguf`	Q5_K_S	5.5 GB	5.49	Higher quality.
`LFM2.5-8B-A1B-Uncensored-Q5_K_M.gguf`	Q5_K_M	5.7 GB	5.69	Higher quality, marginally larger.
`LFM2.5-8B-A1B-Uncensored-Q6_K.gguf`	Q6_K	6.5 GB	6.56	Near-lossless.
`LFM2.5-8B-A1B-Uncensored-Q8_0.gguf`	Q8_0	8.4 GB	8.50	Effectively lossless vs the BF16 source.
`LFM2.5-8B-A1B-Uncensored-BF16.gguf`	BF16	16 GB	16.00	Full precision, identical numerics to the source HF model.

Not sure which to pick? Start with Q4_K_M. Go up to Q5/Q6/Q8 if you have the memory and want maximum fidelity; drop to Q3 or Q2 only if you're memory-constrained. Because this is an MoE with only ~1B active parameters per token, inference throughput is fast even at the larger quants if your hardware has the RAM.

Usage

llama.cpp (auto-download from this repo)

bash

# Interactive chat — downloads the chosen quant automatically
llama-cli -hf zaakirio/LFM2.5-8B-A1B-Uncensored-GGUF:Q4_K_M

# OpenAI-compatible server
llama-server -hf zaakirio/LFM2.5-8B-A1B-Uncensored-GGUF:Q4_K_M -c 4096

Or, with a file you've already downloaded:

llama-cli -m LFM2.5-8B-A1B-Uncensored-Q4_K_M.gguf -p "Hello, who are you?"

Ollama

ollama run hf.co/zaakirio/LFM2.5-8B-A1B-Uncensored-GGUF:Q4_K_M

LM Studio / Jan

Search for zaakirio/LFM2.5-8B-A1B-Uncensored-GGUF in the in-app model browser, or download a .gguf file from this page and load it.

Download a single file

bash

pip install -U "huggingface_hub[cli]"
hf download zaakirio/LFM2.5-8B-A1B-Uncensored-GGUF \
  --include "LFM2.5-8B-A1B-Uncensored-Q4_K_M.gguf" --local-dir ./

Prompt format

The chat template is embedded in the GGUF files, so chat-aware tools apply it automatically. For reference, it is ChatML-style:

<|startoftext|><|im_start|>system
{system_prompt}<|im_end|>
<|im_start|>user
{prompt}<|im_end|>
<|im_start|>assistant

About the base model

This is a decensored derivative produced with Heretic (automatic directional ablation). Compared with the original LFM2.5-8B-A1B:

Metric	Decensored	Original
Refusals (/100 harmful prompts)	0	0
KL divergence (harmless prompts)	0.0481	0 (by definition)

The base LFM2.5-8B-A1B measured 0–2 / 100 refusals on Heretic's marker-based detector (compared to ~98 / 100 for its smaller sibling), suggesting it is comparatively compliant out of the box. The abliteration still makes real, measurable changes to the attention and dense MLP projections (KL ≈ 0.05).

See the source model card for the full abliteration parameters and run details.

Intended use & disclaimer

This model has had its refusal behavior substantially removed and will comply with requests the original model would have declined. It is provided for research and unrestricted local use. You are responsible for how you use it and for complying with all applicable laws and with the base model's lfm1.0 license, which carries over to this derivative.

Provenance

Quantized from zaakirio/LFM2.5-8B-A1B-Uncensored (BF16) using llama.cpp convert_hf_to_gguf.py + llama-quantize.
Base model: LiquidAI/LFM2.5-8B-A1B
Decensoring tool: Heretic by p-e-w