Hacker News
IBM CEO says open weight models are now irresistibly good and cheap on-prem
AIFirst seen 130 points80 comments
IBM's CEO told investors that open weight AI models are now good enough and cheap enough for companies that run AI on their own computers. He said they can do most everyday office tasks at a fraction of the cost of top closed models.
What happened
IBM CEO Arvind Krishna spoke on the company's Q3 earnings call. He said open weight models now make strong business sense for on-prem use. On-prem means the company runs the software on its own computers instead of in the cloud. Two years ago, the best closed models were clearly better than open ones. Now the gap has shrunk a lot.
Key numbers
- In a broad test, a top model might score 70, while an open model scored 20 or 30 in 2024.
- By late 2025, the gap had shrunk to about 70 versus 50.
- Open weight models cost between 25 percent and 10 percent of the token cost of top frontier models. A token is a small piece of text that AI models count when they charge for use.
- Krishna says open models are about 90 percent as effective as the best models.
- IBM's infrastructure revenue rose 5 percent to $3.7 billion. Software revenue rose 9 percent to $7.8 billion.
Why the gap is closing
Open weight models share their trained settings so anyone can download and run them. Closed models are kept private by their makers. Krishna says that for most tasks, such as summarizing an email, checking a contract, or answering a customer call, a cheaper model works just as well. Only a few tasks need the very best model.
Why it matters
Many companies want AI but worry about cost and data privacy. If cheaper models work well, more firms may run AI inside their own walls. Krishna also warned that running these systems is hard. Companies need skilled engineers to set up networks, group computers together, and store results so answers come back fast. IBM plans to sell software to make this easier and expects more sales of its Spyre accelerator cards and its z17 mainframes.
What to watch
- Whether businesses really move work to cheaper open models.
- How much IBM's new rackable systems and Spyre cards sell.
- Whether IBM's software makes on-prem AI easier to set up.
- Whether Red Hat's virtualization deals grow past the $800 million already won.
Glossary
- open weight models
- AI models whose trained settings are shared publicly, so anyone can download and run them.
- on-prem
- Software that runs on a company's own computers instead of in a cloud service.
- Red Hat
- An IBM company that sells Linux and cloud software for businesses.
- Spyre
- IBM's mainframe accelerator card that speeds up AI work on its mainframe computers.
- z17
- IBM's current mainframe computer line, used by large companies for heavy data work.
- token
- A small piece of text, such as a word or part of a word, that AI models count when they charge for use.