Russian Researchers Compress LLMs with HARP Algorithm: 2 Bits and a 2.1x Speedup

⚡ The gist in 5 seconds - Russian researchers have presented HARP, an algorithm for compressing large language models. - In tests, the 2-bit version of the model ran 2.1x faster than the uncompressed one. - Even under heavy compression, a significant portion of the original model's capabilities is preserved. ### 🔍 What was found Russian researchers have developed HARP, an algorithm designed for aggressive compression of large language models. The method makes it possible to convert LLMs to a 2-bit format — that is, to drastically reduce the bit width of the weights. According to test results, the 2-bit version of the model ran 2.1x faster than the uncompressed one while, as reported, retaining a significant portion of its capabilities. The mechanics of HARP are only disclosed in the source at the level of the claimed result: it is a weight compression algorithm, not a standalone model. Specific architectural details, benchmark sets, and the list of tested models are not provided in the source — only the final "2 bits vs. full precision" ratio and the measured 2.1x speedup are known. ### 💡 Why it matters 2-bit quantization is one of the most aggressive compression regimes: it is at such bit widths that models typically lose coherence in their responses, so the claimed preservation of capabilities is the main...