IBM has released Granite 4.2, a new family of open-weight language models aimed at workloads that need long context, tool use, and flexible compute settings. The models are available in 3B, 8B, and 30B sizes, giving developers several options depending on the task and resource constraints.
The release also includes Granite Speech 5.0 Turbo CTC models for transcription. Together, the language and speech models expand IBM's Granite lineup with systems that are available under the Apache 2.0 license on Hugging Face, Ollama, GitHub, and other platforms.
What IBM Released With Granite 4.2
The Granite 4.2 language models were trained from scratch on about 15 trillion tokens, according to IBM. They support context windows up to 512,000 tokens, which gives them room to process very large inputs in a single interaction.
IBM is releasing the family in three model sizes: 3B, 8B, and 30B. That range matters because the smallest model can target lighter deployments, while the larger variants are positioned for tasks that need more capability.
All Granite 4.2 models support OpenAI-format tool calling. They also run on vLLM or SGLang, giving teams more than one serving path for deployment.
The models can switch between "thinking" and "non-thinking" modes. In practical terms, this gives users a way to match the amount of compute to the difficulty of the task. IBM also describes a "low-effort" mode for saving resources on simple queries.
Agentic RL Is The Key Addition
The 8B and 30B Granite 4.2 variants receive what IBM calls "agentic RL" training. In that process, the models learn in real sandbox environments where they use tools, write and run code, and search the web.
This is the clearest signal of what IBM wants the larger Granite 4.2 models to handle. Instead of only producing text responses, the models are trained around actions that modern AI systems increasingly need to perform:
- Using external tools during a task
- Writing code as part of a workflow
- Running code in an environment
- Searching the web when a task requires outside information
The OpenAI-format tool calling support is also important for compatibility. It means developers working with tool-based AI workflows have a familiar interface pattern available across all Granite 4.2 models.
The distinction between the 3B model and the 8B and 30B variants is worth noting. The source describes agentic RL training only for the 8B and 30B models, while all models share support for OpenAI-format tool calling and operation on vLLM or SGLang.
Long Context And Compute Control
A context window up to 512,000 tokens allows Granite 4.2 models to work with large volumes of input at once. For users, the practical implication is that the model can be given more material before it has to respond, summarize, analyze, or call tools.
The ability to toggle between "thinking" and "non-thinking" modes gives another layer of control. Not every prompt needs the same amount of reasoning effort, and IBM's framing suggests users can choose a more efficient setting when the query is simple.
The "low-effort" mode fits that same idea. If a request is straightforward, spending fewer resources can be more appropriate than using heavier compute by default. For developers and teams running repeated tasks, that distinction can affect how they design their workflows.
These settings make Granite 4.2 less of a one-speed system. The family is presented as a set of models that can vary both by size and by compute behavior, depending on the task being handled.
Granite Speech 5.0 Turbo CTC Expands The Release
IBM also introduced Granite Speech 5.0 Turbo CTC models. These models have just 470 million parameters, a much smaller scale than the Granite 4.2 language models described in the same release.
According to IBM, the new speech models are twice as fast as the previous leaders on the Open ASR Leaderboard. IBM also says they can transcribe three hours of audio in one second.
That gives the release a broader scope than language generation alone. The Granite update covers both text-focused models and speech transcription, with the speech models emphasizing speed and compact size.
Why The Apache 2.0 Release Matters
All of the models are available under the Apache 2.0 license. The source lists Hugging Face, Ollama, GitHub, and other platforms as distribution points, which makes the models accessible through channels developers already use.
The combination of open weights, tool calling, long context, and agentic training defines the Granite 4.2 announcement. IBM is not only releasing another set of language models by size; it is putting forward models designed to operate inside workflows where tools, code execution, web search, and compute control all matter.
For teams evaluating open-weight AI systems, the release adds another option with clearly stated model sizes, deployment routes, context limits, and licensing terms. The most capable agentic features described in the source apply to the 8B and 30B variants, while the full Granite 4.2 family shares the broader open-weight release model.