Try LFM • Docs • LEAP • Discord

LFM2.5-350M-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/LFM2.5-350M

🏃 How to run LFM2.5

Example usage with llama.cpp:

llama-cli -hf LiquidAI/LFM2.5-350M-GGUF --conversation \
    --temp 0.1 --top-k 50 --repeat-penalty 1.05

QAD Q4_0 GGUF

The Quantization-Aware Distillation (QAD) checkpoint is available as LFM2.5-350M-QAD-Q4_0.gguf.

This is distinct from the post-training-quantized LFM2.5-350M-Q4_0.gguf; both use the GGUF Q4_0 format.

Example usage with llama.cpp:

llama-cli -hf LiquidAI/LFM2.5-350M-GGUF \
  --hf-file LFM2.5-350M-QAD-Q4_0.gguf \
  -p "What is C. elegans?"

QAD source weights (safetensors)

The original FP32 QAD source checkpoint is available in qad/, with its model config, tokenizer, generation defaults, and the same chat template as the released QAD GGUF. It can be loaded in Transformers by passing subfolder="qad":

from transformers import AutoModelForCausalLM, AutoTokenizer

repo_id = "LiquidAI/LFM2.5-350M-GGUF"
tokenizer = AutoTokenizer.from_pretrained(repo_id, subfolder="qad")
model = AutoModelForCausalLM.from_pretrained(
    repo_id, subfolder="qad", dtype="auto", device_map="auto"
)

inputs = tokenizer.apply_chat_template(
    [{"role": "user", "content": "What is 2 + 2?"}],
    tokenize=True, add_generation_prompt=True, return_dict=True, return_tensors="pt",
).to(model.device)
outputs = model.generate(**inputs, max_new_tokens=128)
print(tokenizer.decode(outputs[0, inputs["input_ids"].shape[-1]:], skip_special_tokens=True))

These weights are intended for fine-tuning and experimentation. Published QAD results apply to the Q4_0 GGUF; direct FP32/BF16 inference and other quantization formats may behave differently. See the source checkpoint documentation for validation details and the license.