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

Krishna put open-weight models at about 90 percent of frontier effectiveness, priced at 10 to 25 percent of frontier token cost. In 2024, a top frontier model might score 70 on a broad benchmark while an open model scored only 20 to 30; by late 2025, the open model's score rose to about 50 against the same 70. He estimated that an average large enterprise could begin running open-weight models with perhaps two or three racks of infrastructure. Krishna also said buyers will usually pick the cheaper option, since cheaper alternative models are adequate for most routine enterprise work, including email summaries, customer calls and contract reviews. Krishna warned that deployment requires specialized engineering talent. Commenters are split between viewing open weight models as genuinely good enough now for local/on-prem use and dismissing the CEO's claims as self-serving hardware promotion. For