onebitllms is a lightweight python package that can be used to easily perform Large-Language Model (LLMs) fine-tuning following the training procedure from BitNet, to produce customized 1.58-bit LLMs.

Requirements

Currently the library only works for NVIDIA CUDA compiled GPUs.

Installation

Download the package through pip:

pip install onebitllms

or directly from source:

pip install git+https://github.com/tiiuae/onebitllms.git

Getting started

1.58bit Fine-tuning

Simply use the replace_linear_with_bitnet_linear after loading the pre-quantized checkpoint, and use directly that model for fine-tuning (e.g. with SFTTrainer from trl library):

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from onebitllms import replace_linear_with_bitnet_linear

model_id = "tiiuae/Falcon-E-1B-Base"
revision = "prequantized"

tokenizer = AutoTokenizer.from_pretrained(
    model_id, revision=revision
)

model = AutoModelForCausalLM.from_pretrained(
    model_id,
    revision=revision,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

model = replace_linear_with_bitnet_linear(model)

# Do the training here ...

After training your model, make sure to quantize the final checkpoint in 1-bit precision with the method quantize_to_1bit:

from onebitllms import quantize_to_1bit

model_output_dir = ""
quantized_model_output_dir = ""

quantize_to_1bit(model_output_dir, quantized_model_output_dir)

Example scripts can be found in examples/. We use the following command on a 8xA10G 23GB instance and training took ~8.5 hours.

python examples/sft.py \
    --model_name_or_path tiiuae/Falcon-E-1B-Base \
    --model_revision prequantized \
    --torch_dtype bfloat16 \
    --learning_rate 0.0001 \
    --dataset_name trl-lib/Capybara \
    --per_device_train_batch_size 1 \
    --output_dir Falcon-E-Capybara \
    --logging_steps 1 \
    --save_strategy steps \
    --save_steps 100 \
    --packing \
    --gradient_accumulation_steps 16

Once your 1.58bit model is ready, we suggest you to deploy it with bitnet.cpp package.

Using quantization triton kernels

You can also inject the BitNetLinear classes into your pre-training framework:

from onebitllms import BitNetLinear

# inject it in your model classes for pre-training ..

You can also use the triton kernels directly for a more fine-grained control over the operations:

from onebitllms import activation_quant_triton, weight_quant_triton

Revert back to bfloat16 format

From our experiments, the BitNet checkpoints are universal, meaning we can revert back to bfloat16 format with minimal performance degradation. You can use the method convert_to_bf16 after training your model:

from onebitllms import convert_to_bf16

model_output_dir = ""
quantized_model_output_dir = ""

convert_to_bf16(model_output_dir, quantized_model_output_dir)

Common questions

What is 1-bit fine-tuning?

We term 1-bit fine-tuning as simply doing continuous training from a pre-quantized BitNet compatible checkpoint. As of today, there are multiple ongoing work that explores fine-tuning existing checkpoints into BitNet format but this often leads to poor performance.

What models can I fine-tune?

To the best of our knowledge, as of today, only Falcon-Edge and recent Microsoft BitNet series models published their pre-quantized checkpoints. If in the future other models gets published together with their pre-quantized checkpoints, they should be compatible with onebitllms out-of-the-box.

What else can I do with onebitllms?

You can also use the BitnetLinear class exposed in this package and use it inside your pre-training / fine-tuning framework. In contrary to existing implementation, we use optimized triton kernels for computing the quantization errors making the pre-training and fine-tuning much faster than existing implementations. From our experiments, we estimate the overheads between non-BitNet and BitNet pre-training to be around ~20% (to be confirmed with more rigourous analysis).

Is LoRA supported in onebitllms?

LoRA is not supported with this package and remains a very interesting research question. Unlocking LoRA with onebitllms could open-up exciting opportunities such as being able to fine-tune a BitNet 7B on a free-tier Google Colab instance.

Can onebitllms used for inference?

As of today, BitNet models are extremely interesting for CPU deployment. We strongly encourage users to deploy their models with bitnet.cpp package after fine-tuning it with onebitllms. If you want to run it on GPU, to the best of our knowledge, you can use HuggingFace's transformers native integration of BitNet models.

Can I contribute to onebitllms?

Of course. Contributions to enhance the codebase, introduce new features and example scripts are strongly encouraged.

Citation

If you find this work useful for your research and work, please consider citing us, as well as citing the foundational work behind BitNet models:

@misc{tiionebitllms,
    title = {Falcon-E, a series of powerful, universal and fine-tunable 1.58bit language models.},
    author = {Falcon-LLM Team},
    month = {May},
    url={https://github.com/tiiuae/onebitllms}, 
    year = {2025}
}

@misc{wang2025bitnetcppefficientedgeinference,
      title={Bitnet.cpp: Efficient Edge Inference for Ternary LLMs}, 
      author={Jinheng Wang and Hansong Zhou and Ting Song and Shijie Cao and Yan Xia and Ting Cao and Jianyu Wei and Shuming Ma and Hongyu Wang and Furu Wei},
      year={2025},
      eprint={2502.11880},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2502.11880}, 
}

@misc{,
      title={1.58-Bit LLM: A New Era of Extreme Quantization}, 
      author={Mohamed Mekkouri and Marc Sun and Leandro von Werra and Thomas Wolf},
      year={2024},
}

@misc{ma2024era1bitllmslarge,
      title={The Era of 1-bit LLMs: All Large Language Models are in 1.58 Bits}, 
      author={Shuming Ma and Hongyu Wang and Lingxiao Ma and Lei Wang and Wenhui Wang and Shaohan Huang and Li Dong and Ruiping Wang and Jilong Xue and Furu Wei},
      year={2024},
      eprint={2402.17764},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2402.17764}, 
}

@misc{wang2023bitnetscaling1bittransformers,
      title={BitNet: Scaling 1-bit Transformers for Large Language Models}, 
      author={Hongyu Wang and Shuming Ma and Li Dong and Shaohan Huang and Huaijie Wang and Lingxiao Ma and Fan Yang and Ruiping Wang and Yi Wu and Furu Wei},
      year={2023},
      eprint={2310.11453},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2310.11453}, 
}