Bitnet B 1.58 (Documentation: https://huggingface.co/microsoft/bitnet-b1.58-2B-4T and chatbot on Huggingface: <a href="https://huggingface.co/spaces/suayptalha/Chat-with-Bitnet-b1.58-2B-4T”>https://huggingface.co/spaces/suayptalha/Chat-with-Bitnet-b1.58-2B-4T) is a Microsoft language model open-source that works with 1 bit per parameter (compared to 16 or 32 for other equivalent models).

By this reduction:

• Bitnet B 1.58 can run directly with a computer’s processor without a dedicated graphics card,
• It occupies 0.4 GB of memory compared to 2 to 5 GB for equivalent models,
• Despite a slight loss of precision (in the tests, what approximations in French and difficulties in calculating the durations), it offers answers of a quality close to standard models,
• finally, it consumes less power compared to other comparable LLMs (see the technical report on Arxiv: <a href="https://arxiv.org/html/2504.12285v1″>https://arxiv.org/HTML/2504.12285v1 and screenshot at the end of the article)

In the Huggingface chatbot, there are several options: given “role”, number of tokens and temperature.

Source: <a href="https://arxiv.org/html/2504.12285v1″>https://arxiv.org/HTML/2504.12285v1