BitNet
In plain wordsA family of AI language models from Microsoft in which most of the learned numbers are limited to minus one, zero or one, so they need far less memory and energy to run.
A family of transformertransformer, n.: The design behind today's chatbots and much of modern AI. It handles text by letting every word look at every other word and decide which of them matter for understanding it. language modelslarge language model, n.: A computer program trained on enormous amounts of text to predict the words that come next. At vast scale, that simple skill lets it chat, answer questions, translate, summarize and write code. and the architecture behind them, from Microsoft Research, in which the weights of the linear layers are constrained to extremely low precision during training itself rather than quantized afterwards. The original BitNet (2023) replaced standard linear layers with BitLinear layers holding binary (±1) weights and 8-bit activations, trained with full-precision latent weights and straight-through gradients. BitNet b1.58 (2024) made every weight ternary, one of {−1, 0, +1}, which carries log₂3 ≈ 1.58 bits; its authors reported matching full-precision models of the same size and training data, while matrix multiplication reduces mostly to additions. BitNet b1.58 2B4T (April 2025), with about two billion parameters trained on four trillion tokens, was released as open weights with GPU and CPU inference code. Realizing the gains requires kernels or hardware designed for ternary arithmetic.
The BitNet model's weights take about 0.4 GB, so it runs on a laptop CPU at reading speed.
Origin
Introduced by Hongyu Wang, Shuming Ma, Furu Wei and colleagues at Microsoft Research and partner universities in October 2023. 'The Era of 1-bit LLMs' (February 2024) presented the ternary b1.58 variant and popularized the phrase '1-bit LLM', though its weights are strictly 1.58-bit.
Usage note
Native low-bit training differs from post-training quantizationpost-training quantization, n.: Shrinking an already-trained AI model by storing its numbers in fewer bits, without training it again; a small sample of example inputs is usually enough to fine-tune the rounding.: the model learns under the constraint from the start, avoiding the accuracy collapse PTQ suffers at very low bit widths. Whether such models stay competitive at frontier scale is unresolved as of October 2026.
Related
- quantizationquantization, n.: Storing an AI model's numbers with fewer digits, like rounding prices to the nearest dollar, so the model takes less memory and runs faster, usually with only a small loss of accuracy.
- quantization-aware trainingquantization-aware training, n.: Training or fine-tuning an AI model while simulating the low-precision numbers it will use later, so that it learns to work well despite the rounding.
- post-training quantizationpost-training quantization, n.: Shrinking an already-trained AI model by storing its numbers in fewer bits, without training it again; a small sample of example inputs is usually enough to fine-tune the rounding.
- on-device AIon-device AI, n.: AI that runs directly on your own phone, laptop, car or headset rather than on a company's distant servers, so it can work offline, respond instantly and keep your data on the device.
- model compressionmodel compression, n.: The family of techniques for making AI models smaller, faster or cheaper to run while keeping almost all of their accuracy.
Sources
- Wang, H. et al. (2023). BitNet: Scaling 1-bit Transformers for Large Language Models. arXiv:2310.11453. arxiv.org/abs/2310.11453
- Ma, S. et al. (2024). The Era of 1-bit LLMs: All Large Language Models are in 1.58 Bits. arXiv:2402.17764. arxiv.org/abs/2402.17764
- Ma, S. et al. (2025). BitNet b1.58 2B4T Technical Report. arXiv:2504.12285. arxiv.org/abs/2504.12285
Mentioned in
quantizationquantization, n.: Storing an AI model's numbers with fewer digits, like rounding prices to the nearest dollar, so the model takes less memory and runs faster, usually with only a small loss of accuracy., quantization-aware trainingquantization-aware training, n.: Training or fine-tuning an AI model while simulating the low-precision numbers it will use later, so that it learns to work well despite the rounding., straight-through estimatorstraight-through estimator, n.: A shortcut that lets networks learn even when they contain steps with no useful slope, such as rounding: during learning, the computer simply pretends the step passed its input straight through.