IBM expanded its Granite model line with downloadable reasoning-focused models aimed at predictable self-hosting, including agent training in the two larger variants. The open-weight family includes three-, eight- and thirty-billion-parameter versions with a native 128,000-token context window. Granite 4.2 is available in three-, eight- and thirty-billion-parameter variants. All three models have a 128,000-token context window.
The eight- and thirty-billion-parameter versions received specialized agentic reinforcement training. The models are open weight and can be downloaded for self-hosting. IBM describes the release as reasoning-focused. These reported details define the immediate scale, timing and institutions involved without treating preliminary statements as final findings.
Local models can avoid per-token cloud fees but require hardware, deployment and evaluation work. Reasoning modes can improve multi-step performance while increasing latency and compute use. Model routers are increasingly used to send tasks to differently sized systems. The chronology separates the new development from older background and keeps official descriptions attached to the officials or organizations that made them.
The announcement and early reporting did not establish performance across every workload or the total cost of a production deployment. That uncertainty remains part of the report because later court orders, technical examinations, audited data, investigative records or field access could change the account.
The checked reporting on IBM Releases Granite 4.2 Models for Local and Enterprise Deployment identifies two concrete follow-up points: independent benchmark and safety evaluations and enterprise reports comparing local and hosted operating costs. Neither had produced a later attributable outcome before this edition's research cutoff.
The source record also distinguishes the confirmed event—Granite 4.2 is available in three-, eight- and thirty-billion-parameter variants.—from the reported consequence—The models are open weight and can be downloaded for self-hosting.—and from the unresolved boundary: The announcement and early reporting did not establish performance across every workload or the total cost of a production deployment. This avoids assigning motive, cause or certainty beyond the cited reporting.
Ars Technica supplied the dated reporting and direct statements used for IBM Releases Granite 4.2 Models for Local and Enterprise Deployment. Where a government, company, military, hospital or advocacy participant described its own conduct, this article treats that description as an attributed claim rather than independent verification.