Open4bits / LFM2.5-1.2B-Base-Quantized

This repository provides multiple quantized variants of the LFM 2.5 Base (1.2B parameters) model for efficient inference and deployment.

The original model is developed and released by LiquidAI:

Original model:
https://huggingface.co/LiquidAI/LFM2.5-1.2B-Base

These quantizations are maintained and published by ArkAiLab under the Open4bits organization to improve accessibility across a wide range of hardware.

---

Available Quantization Formats

Each format is stored in a separate directory:

• FP16 – Baseline half-precision weights
• FP8 – High-performance low-precision format (GPU support required)
• INT8 – Balanced performance and memory usage (BitsAndBytes)
• NF4 (4-bit) – Maximum compression using BitsAndBytes double quant

---

Model Information

• Model Name: LFM 2.5 Base
• Parameters: ~1.2B
• Architecture: Custom LiquidAI architecture
• Original Author: LiquidAI
• Quantized By: ArkAiLab (Open4bits)

This model requires trust_remote_code=True when loading.

---

Quantization Details

• Quantized using PyTorch and Hugging Face Transformers
• INT8 and NF4 formats use BitsAndBytes
• FP8 provided where hardware support allows
• No GPTQ, AWQ, or llama.cpp used
• Safe for Google Colab and Kaggle

---

Usage Example

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Open4bits/LFM2.5-1.2B-Base-Quantized"

tokenizer = AutoTokenizer.from_pretrained(
    model_id,
    trust_remote_code=True
)

model = AutoModelForCausalLM.from_pretrained(
    model_id,
    trust_remote_code=True,
    device_map="auto"
)

inputs = tokenizer("Hello, world!", return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=50)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

---

Organization

This repository is maintained by ArkAiLab under the Open4bits initiative.

ArkAiLab (Main Organization): https://huggingface.co/ArkAiLab-Adl

Open4bits (Quantization Projects): https://huggingface.co/Open4bits

---

License

This repository follows the same license as the original LiquidAI model.

Please refer to the original model repository for full licensing details.

---

Disclaimer

This is an unofficial quantized release.

All credit for the original model architecture and training goes to LiquidAI.