LiquidAI/LFM2.5-8B-A1B-GGUF

🤗 Hugging Face sourcetext-generationlfm1.01B activated66 GBGGUF✓ 7 checksumsupdated 52d ago

Try LFM • Docs • LEAP • Discord

LFM2.5-8B-A1B-GGUF

LFM2 is a new generation of hybrid models developed by Liquid AI, specifically designed for edge AI and on-device deployment. It sets a new standard in terms of quality, speed, and memory efficiency.

Find more details in the original model card: https://huggingface.co/LiquidAI/LFM2.5-8B-A1B

🏃 How to run LFM2

Example usage with llama.cpp:

llama-cli -hf LiquidAI/LFM2.5-8B-A1B-GGUF --conversation \
    --temp 0.2 --top-k 80 --repeat-penalty 1.05

@LiquidAI is unclaimedSign in with Hugging Face to claim this handle.

Claim @LiquidAI

Seed this model

Have the files? Submit a torrent.

Submit a torrent →

Package the files with one command

Have the complete files and permission to share them? Replace the example folder path. The script checks the files, makes a torrent and submits its magnet and checksums. Keep your torrent client seeding. Get your community key from your account, and read the script before running it.

curl -fsSL https://pirateface.co/package.sh | bash -s -- --repo LiquidAI/LFM2.5-8B-A1B-GGUF ./model-folder

Files and sources

SourceAvailable on Hugging Face

Checksums7 recorded

TorrentNot listed

SwarmNot checked

Source revision: 49c14831707011e64d70b2ebd8462ba08d608434
Witnessed: 14 Oct 2026, 14:11 UTC
Weight files: 7
Model size: 66 GB
Downloads · last month: 245,624
Publisher: Unclaimed

View provenance entry #30757 →

Speed on real hardware

LocalMaxxing

Output speed for one request at a time, measured by people running this model on their own machines.

RTX 5070llama.cpp · Q6_K · 1 run350 tok/s

llama-bench -m LFM2.5-8B-A1B-Q6_K.gguf -p 512 -n 128 -b 2048 -ub 512 -ngl -1 -o json -r 1

View run on LocalMaxxing

RTX 3080llama.cpp · Q6_K · 2 runs278 tok/s

llama-server \
  -m LFM2.5-8B-A1B-Q6_K.gguf \
  --host 127.0.0.1 --port 8099 \
  --device CUDA0 --n-gpu-layers 99 \
  --ctx-size 128000 \
  --flash-attn on \
  --cache-type-k q8_0 --cache-type-v q8_0 \
  --parallel 1 --jinja \
  --spec-type none

View run on LocalMaxxing

2× RTX 3090llama.cpp · Q8_0 · 4 runs259 tok/s
```
llama-server -m LFM2.5-8B-A1B-Q8_0.gguf -ngl 999 -fa1 -c 0
```
View run on LocalMaxxing

2× RTX 3060llama.cpp · Q4_K_M · 7 runs218 tok/s

llama-bench -m LFM2.5-8B-A1B-Q4_K_M.gguf -p 4096 -n 128 -ngl 24 -sm layer -fa on -ctk f16 -ctv f16 -r 5

View run on LocalMaxxing

2× RX 570 4GBllama.cpp · Q4_K_M · 4 runs160 tok/s

# Remote endpoint: http://192.168.1.182:8085  servedModel: LFM2.5-8B-Unsloth

View run on LocalMaxxing

RX 570 4GBllama.cpp · Q4_K_M · 2 runs74.7 tok/s

llama-server --model LFM2.5-8B-Unsloth --n-gpu-layers 32 --ctx-size 32768 --flash-attn on --threads 18 --threads-batch 18 --numa numactl --split-mode layer --parallel 1 --cache-type-k q4_0 --cache-type-v q4_0

View run on LocalMaxxing

Median of 21 community runs on 7 setups, from LocalMaxxing.

Run it yourself: setup guide