Tech giant IBM (IBM) has launched ZipNN, a free and open-source tool that compresses AI models to reduce storage costs and speed up transfers without lowering performance. Unlike older methods that remove or change parts of a model, ZipNN uses lossless compression, which shrinks files by removing repeated patterns and restores them exactly when needed. This technique, similar to ZIP files on your computer, can lower storage costs by 33% and boost transfer speeds by up to 150%, which is a big deal for companies and researchers who need to store or move large AI models.
Don’t Miss TipRanks’ Half-Year Sale
The idea for ZipNN came when IBM researchers, working with universities like MIT and Tel Aviv University, discovered that part of a model’s data—called exponents—follows a predictable pattern. These exponents are part of floating-point numbers used to store information in AI models. Out of 256 possible values, just 12 make up nearly all of the data. Interestingly, ZipNN separates these predictable exponents from the more random parts and uses a technique called Huffman encoding to compress them efficiently. This method works especially well for modern BF16-format models like Meta’s Llama or IBM’s Granite. In fact, it reduced their size by 33% and outperformed Meta’s Zstandard (zstd) by 11%.
ZipNN also includes a technique called “byte grouping,” which splits the data into smaller parts to identify even more patterns. In one case, it cut the size of Meta’s xlm-RoBERTa model by over half. In addition, Hugging Face, a major AI platform, has already adopted this method and reported saving 20% on storage costs. Looking ahead, ZipNN could help reduce the cost of saving unfinished “checkpoint” versions of AI models and speed up downloads for millions of users.
What Is the Target Price for IBM?
Turning to Wall Street, analysts have a Moderate Buy consensus rating on IBM stock based on seven Buys, five Holds, and two Sells assigned in the past three months, as indicated by the graphic below. Furthermore, the average IBM price target of $269.46 per share implies 7.8% downside risk.
See more IBM analyst ratings
Disclaimer & DisclosureReport an Issue