IBM has released its Granite 4.2 language model family in 3B, 8B, and 30B sizes, The Decoder reports, citing IBM. The models are open-weight and available under the Apache 2.0 license on Hugging Face, Ollama, GitHub, and other platforms, according to the report. The technical pitch is a mix of longer context, controllable inference behavior, and tool-use training. The Decoder reports that Granite 4.2 was trained from scratch on about 15 trillion tokens and supports context windows of up to 512,000 tokens. IBM says the models can switch between “thinking” and “non-thinking” modes, with a low-effort mode meant to conserve compute on simpler tasks. The agentic component is not uniform across the whole family. According to The Decoder, IBM applied what it calls “agentic RL” training to the 8B and 30B variants, teaching them to use tools, write and run code, and search the web in real sandbox environments. The report says all models support OpenAI-format tool calling and can run on vLLM or SGLang. That makes Granite 4.2 relevant to teams evaluating deployable open-weight models for agent workflows, not just chat completion. The combination of OpenAI-format tool calling and support for common serving stacks gives builders a clearer path to test the models inside existing inference and orchestration setups, assuming the reported compatibility holds in their environments. The Decoder also reports a related speech-model update: IBM’s Granite Speech 5.0 Turbo CTC models have 470 million parameters, and IBM says they are twice as fast as the previous leaders on the Open ASR Leaderboard. IBM also claims they can transcribe three hours of audio in one second. Those performance figures are IBM-reported in the supplied material and should be treated as claims to benchmark against real workloads. The Hugging Face item in this cluster is an IBM Granite post titled “Granite 4.2 LLMs: How They’re Built,” but the supplied feed item includes no summary. That supports the presence of a Granite 4.2 technical post on Hugging Face, while the specific numbers and deployment details above come from The Decoder’s report. Who benefits: Teams that prefer permissively licensed open-weight models benefit from another IBM model family to evaluate. Developers already using Hugging Face, Ollama, GitHub, vLLM, or SGLang may have a lower integration path if the reported support matches their stack. Who's exposed: Organizations that treat vendor-reported speed, context, or agentic claims as production-ready evidence are exposed to execution risk. The supplied material does not include independent benchmark validation for the language models.