Models

Available Models

BitNet supports a variety of 1-bit quantized models. These models are pre-quantized and ready for inference. For usage instructions, see our Usage Guide. For installation instructions, check our Installation Guide.

BitNet-b1.58 Models

BitNet-b1.58 is a native BitNet architecture using 1.58-bit quantization (ternary-like with -1, 0, +1).

Falcon3 Models

Falcon3 models quantized to 1-bit using BitNet quantization techniques. These are instruction-tuned models suitable for conversational AI applications.

Llama3 Models

Llama3-based models with BitNet quantization, trained on large token counts.

Downloading Models

Models can be downloaded directly from HuggingFace using the HuggingFace CLI:

huggingface-cli download microsoft/BitNet-b1.58-2B-4T-gguf \
  --local-dir models/BitNet-b1.58-2B-4T

For .safetensors models, you'll need to convert them to GGUF format. See our Usage Guide for conversion instructions.

Model Selection Guide

Choose by Use Case

• General Purpose Text Generation: BitNet-b1.58-2B-4T or BitNet-b1.58-3B
• Conversational AI / Chat: Falcon3-Instruct models (1B, 3B, 7B, or 10B depending on requirements)
• Research & Experimentation: Any model depending on research focus
• Resource-Constrained Environments: Smaller models like Falcon3-1B-Instruct or BitNet-b1.58-2B-4T
• Higher Quality Output: Larger models like Falcon3-7B-Instruct or Falcon3-10B-Instruct

Choose by Resource Constraints

• Low Memory (4-8GB): Falcon3-1B-Instruct or BitNet-b1.58-2B-4T
• Medium Memory (8-16GB): Falcon3-3B-Instruct or BitNet-b1.58-3B
• High Memory (16GB+): Falcon3-7B-Instruct, Falcon3-10B-Instruct, or Llama3-8B-1.58

Model Format

BitNet models are distributed in GGUF format, which is optimized for efficient loading and inference. Some models are also available in .safetensors format, which can be converted to GGUF using the provided conversion utilities.

Quantization types available:

• i2_s: 1-bit signed quantization (recommended)
• tl1: Ternary-like quantization variant

Using Models

After downloading a model, you need to set it up before use:

python setup_env.py -md models/BitNet-b1.58-2B-4T -q i2_s

Then run inference as shown in our Usage Guide.

Related Resources

• Usage Guide - How to use models
• Installation Guide - Setup instructions
• Benchmark Guide - Compare model performance
• Documentation - Complete API reference
• Microsoft on HuggingFace - All BitNet models