IBM’s Granite 4.2 family brings open‑weight reasoning models in 3 B, 8 B and 30 B sizes under an Apache 2.0 license, eliminating legal friction for commercial use. The core challenge for developers and enterprises is finding a model that can explain its thinking, adapt to tool use, and run efficiently on existing hardware without costly licensing or complex integration. Granite 4.2 solves this by exposing a simple thinking / low‑effort / non‑thinking switch in the chat template, letting users control the reasoning budget per query. For teams that need agentic behavior—such as code editing, terminal commands, or web search—the 8 B and 30 B versions include a dedicated RL block trained on real sandboxed environments, while the 3 B focuses on foundational alignment.
Deployment is straightforward: the 3 B fits on a laptop via Ollama or LM Studio with GGUF quantizations; the 8 B runs comfortably on a single modern GPU; the 30 B targets enterprises with A100/H100‑class resources or can be served with vLLM using FP8/NVFP4 for higher throughput. All models are decoder‑only dense transformers with 128K‑token context (extendable to 512K in pretraining) and bfloat16 precision, ensuring compatibility with standard inference stacks.
Alongside the language models, IBM released two 470 M‑parameter Granite Speech 5.0 Turbo CTC models that drop the LLM backbone entirely, delivering high‑speed transcription (≈12 600 RTFx on an H200) for contact‑center and media workloads.
In short, Granite 4.2 gives organizations a transparent, controllable, and license‑free path to embed reasoning and agentic capabilities into applications—whether you are a solo developer experimenting on a laptop or a large enterprise scaling AI‑driven DevOps, coding assistants, or speech‑to‑text pipelines.
#AI #MachineLearning #OpenSource #LLM #AgenticAI #SpeechToText