IBM CEO says open weight models are now good and cheap on-prem

Commenters who use open-weight models said they have become capable enough for everyday coding and other work, especially when a person reviews and corrects the output. They cited models such as Kimi, GLM, and DeepSeek as alternatives to frontier systems, and argued that many users do not need the latest benchmark leader. Running models locally can also keep sensitive work away from paid providers and enable offline use. Some described the tools as empowering, while others saw cheaper open models as eroding the frontier labs’ advantage and raising questions about how those labs will sustain enormous investment if older or less costly models are good enough for most tasks.

The divide was over whether self-hosting is actually cheaper and practical. Supporters pointed to lower usage costs, privacy, and greater control; skeptics noted high GPU prices, power costs, setup complexity, and the risk of paying for idle hardware. They said rented cloud compute or model APIs can be more economical and flexible, especially for companies without specialized infrastructure staff. Several commenters also questioned IBM’s claims because IBM sells hardware and services, while others welcomed a major established company backing alternatives to frontier labs. For some, on-premises deployment makes sense when data security or reliable GPU access matters; for others, cloud services remain the better deal.