**Translation**
IBM expanded its open‑weight model family with the new Granite 4.2. These models are offered in three sizes—3 billion, 8 billion, and 30 billion parameters—and are designed for local download and hosting.
Like earlier versions, Granite 4.2 uses a decoder‑only architecture and natively supports a 128 000‑token context window. The 8‑billion and 30‑billion versions include an agent‑based reinforcement‑learning layer that trains them for tasks such as terminal use, web browsing, or employing external tools; the 3‑billion version also supports tools but without the same level of specialized training.
**Reasoning and Explanation**
- The original Persian text describes IBM’s release of Granite 4.2, specifying the three model sizes and their purpose (local download/hosting).
- It notes the decoder‑only architecture and the 128 k token context window, which are direct technical details to be preserved in translation.
- The distinction between the 8 billion/30 billion models (with an agent‑based RL layer for advanced tasks) and the 3 billion model (supports tools but lacks that specialized training) is maintained to reflect the nuance of the source.
- The translation aims for a clear, natural English rendering while keeping the factual content and structure of the original paragraphs.