vkhosla introduces the 1-bit Bonsai 8B model, with memory usage of only 1.15GB.

ME News update. On April 1 (UTC+8), recently, the author vkhosla introduced a 1-bit weight model named 1-bit Bonsai 8B on social media. According to the introduction, the model’s memory footprint is only 1.15 GB, and its intelligence density is claimed to be more than 10 times that of its full-precision corresponding model. The article’s viewpoint states that on edge hardware, the model’s size has been reduced by 14 times, its speed increased by 8 times, and its energy efficiency improved by 5 times, while still remaining competitive compared with other models. The news also mentions that the model has better compressibility, faster speed, smaller size, and higher energy efficiency than the model announced by Google the previous day. The original text does not provide specific performance comparison benchmarks, the publisher, technical implementation details, or complete evaluation results. (Source: InFoQ)