license: other license_name: lfm1.0 license_link: https://huggingface.co/LiquidAI/LFM2.5-24B-A2B/blob/main/LICENSE base_model: pottokao/LFM2.5-24B-A2B-Uncensored pipeline_tag: text-generation tags:

liquid
lfm2.5
moe
uncensored
heretic
ara
gguf
llama.cpp

LFM2.5-24B-A2B-Uncensored-GGUF

GGUF quantizations of pottokao/LFM2.5-24B-A2B-Uncensored (LiquidAI LFM2.5-24B-A2B decensored with Heretic ARA, searched and tested with thinking enabled; see the main repo for results).

File	Size	Notes
`LFM2.5-24B-A2B-Uncensored-Q4_K_M.gguf`	14.5 GB	no imatrix
`LFM2.5-24B-A2B-Uncensored-IQ4_XS.gguf`	12.9 GB	imatrix · full GPU offload on a 16 GB card
`LFM2.5-24B-A2B-Uncensored-Q3_K_M.gguf`	11.5 GB	imatrix · K-quant, better when experts run on CPU (`--n-cpu-moe`)
`LFM2.5-24B-A2B-Uncensored-IQ3_M.gguf`	10.5 GB	imatrix · smallest, plenty of room for long context on 16 GB
`lfm25-uncensored.imatrix`	—	importance matrix used above

IQ quants are fastest fully on GPU; if you offload experts to CPU, prefer the K-quants.

All files keep every attn_output tensor at Q8_0 (--tensor-type attn_output=q8_0): the decensoring edit lives in blk.22.attn_output and blk.26.attn_output, and the 10 attention outputs together are only ~40M parameters.
imatrix computed on the decensored BF16 GGUF with bartowski's calibration_datav3.
Converted and quantized with llama.cpp 0a1ef30.

KV cache is small (only 10 of 40 layers are attention, ~20 KB per token in F16), so most of the VRAM goes to weights.

llama-server -m LFM2.5-24B-A2B-Uncensored-IQ3_M.gguf -ngl 99 -c 32768 --jinja \
  --temp 0.1 --top-k 50 --repeat-penalty 1.05

The model always thinks before answering; allow a generous output budget. Liquid AI's official DSpark draft targets the original weights (acceptance on this model not yet measured).

⚠️ Notice

Refusal behavior has been substantially reduced. Intended for research, red-teaming and personal use; you are responsible for your use and for complying with laws and the base model license.

License & credits

Derivative of LiquidAI/LFM2.5-24B-A2B under the LFM Open License v1.0. Decensoring with Heretic (p-e-w, AGPL-3.0) / ARA. Quantized with llama.cpp.

LFM2.5-24B-A2B-Uncensored-GGUF

GGUF quantizations of pottokao/LFM2.5-24B-A2B-Uncensored (LiquidAI LFM2.5-24B-A2B decensored with Heretic ARA, searched and tested with thinking enabled; see the main repo for results).

File

Size

Notes

LFM2.5-24B-A2B-Uncensored-Q4_K_M.gguf

14.5 GB

no imatrix

LFM2.5-24B-A2B-Uncensored-IQ4_XS.gguf

12.9 GB

imatrix · full GPU offload on a 16 GB card

LFM2.5-24B-A2B-Uncensored-Q3_K_M.gguf

11.5 GB

imatrix · K-quant, better when experts run on CPU (--n-cpu-moe)

LFM2.5-24B-A2B-Uncensored-IQ3_M.gguf

10.5 GB

imatrix · smallest, plenty of room for long context on 16 GB

lfm25-uncensored.imatrix

—

importance matrix used above

IQ quants are fastest fully on GPU; if you offload experts to CPU, prefer the K-quants.

All files keep every attn_output tensor at Q8_0 (--tensor-type attn_output=q8_0): the decensoring edit lives in blk.22.attn_output and blk.26.attn_output, and the 10 attention outputs together are only ~40M parameters.

imatrix computed on the decensored BF16 GGUF with bartowski's calibration_datav3.

Converted and quantized with llama.cpp 0a1ef30.

KV cache is small (only 10 of 40 layers are attention, ~20 KB per token in F16), so most of the VRAM goes to weights.

llama-server -m LFM2.5-24B-A2B-Uncensored-IQ3_M.gguf -ngl 99 -c 32768 --jinja \ --temp 0.1 --top-k 50 --repeat-penalty 1.05

The model always thinks before answering; allow a generous output budget. Liquid AI's official DSpark draft targets the original weights (acceptance on this model not yet measured).