Superpower Daily
IBM Releases Granite 4.2 Open Weights, Reserving Agentic Traini...Harvard Uses 7 AI Faculty Avatars in a $699 Founder Bootcamp, W...Alibaba Raises HK$80B for Full-Stack AI as Investors Question t...MIT’s CrysVCD Steers AI Material Design Toward Stability Before...Anthropic Hires Builder of Google’s First Seven TPUs for In-Hou...LLMScholarBench Tests 22 Models and Finds an Accuracy-Represent...IBM Releases Granite 4.2 Open Weights, Reserving Agentic Traini...Harvard Uses 7 AI Faculty Avatars in a $699 Founder Bootcamp, W...Alibaba Raises HK$80B for Full-Stack AI as Investors Question t...MIT’s CrysVCD Steers AI Material Design Toward Stability Before...Anthropic Hires Builder of Google’s First Seven TPUs for In-Hou...LLMScholarBench Tests 22 Models and Finds an Accuracy-Represent...
Modelspublished

IBM Releases Granite 4.2 Open Weights, Reserving Agentic Training for 8B and 30B

The new family gives self-hosted deployments a choice between a smaller tool-capable model and larger variants trained for code, search and other tool-use tasks.

By 2 min read
IBM Releases Granite 4.2 Open Weights, Reserving Agentic Training for 8B and 30B
Sign in to save and react

Listen to this story

The audio brief

About 1:26
0:001:26
Read transcript
IBM’s new Granite 4.2 family puts a clear capability divide inside its open-weight lineup: the eight-billion- and 30-billion-parameter models received specialized agentic training, while the three-billion model did not. That training covered sandboxed terminal work, code writing and execution, web search, and other external tools. The smaller model can still call tools using the OpenAI format; it simply lacks the larger models’ additional optimization for those tasks. All three models are designed for download and self-hosting, with a native context window of 128,000 tokens. IBM says the family can support up to 512,000 tokens, though that is the stated maximum rather than the standard window. Granite 4.2 also lets organizations choose between thinking, non-thinking, and lower-compute modes, making the reasoning trade-off configurable by task. More reasoning can mean more rigorous answers, but also slower responses and higher compute costs. The models run with vLLM or SGLang and are available under the Apache 2.0 license through Hugging Face, Ollama, GitHub, and other platforms. IBM also released 470-million-parameter Granite Speech 5.0 Turbo CTC models, claiming transcription of three hours of audio per second. The practical question is whether the three-billion model’s basic tool calling is enough for a deployment, or whether agentic workloads justify the larger models and their added compute demands.

Story brief

3 key points

IBM’s Granite 4.2 expands its self-hosted model lineup with 3B, 8B, and 30B variants, but reserves specialized agentic reinforcement-learning training for the two larger models. That training targets terminal work, code execution, web search, and external tools; the 3B model retains tool calling without the same optimization. All models offer a 128,000-token native context, with IBM citing support up to 512,000...

  1. 01

    Only the 8B and 30B models received agentic training in sandboxed terminal, coding, web-search, and tool-use tasks.

  2. 02

    The 3B model supports OpenAI-format tool calling, but lacks the larger variants’ specialized agentic training.

  3. 03

    Granite 4.2 runs with vLLM or SGLang and supports thinking, non-thinking, and lower-compute modes.

IBM has released Granite 4.2, an open-weight language-model family built for download and self-hosting. The central distinction is not just size: its 8B and 30B models receive specialized training for tool use, while the 3B version does not.

The release spans 3B, 8B and 30B parameter variants, each with a native 128,000-token context window. IBM also says Granite 4.2 can support contexts of up to 512,000 tokens—the stated family maximum, rather than the standard window.

Tool support is not the training line

The 8B and 30B variants received agentic reinforcement-learning training for terminal use, web search and external tools. IBM also describes that training as covering code writing and execution, alongside web search, in sandbox environments. The 3B model supports tools, but did not receive the same specialized agentic training.

All Granite 4.2 models support OpenAI-format tool calling and can run on vLLM or SGLang. IBM calls Granite 4.2 the Granite family’s reasoning-focused release; the models can switch between thinking and non-thinking modes, while low-effort mode is intended to conserve resources on simpler queries.

A reasoning-focused release with a compute trade-off

Functional reasoning carries intermediate results through multiple steps. It can produce more rigorous and accurate responses in some cases, but often brings slower response times and higher compute demands. Granite’s selectable modes position that trade-off as a setting organizations can vary by task.

Granite 4.2 is available under the Apache 2.0 license through Hugging Face, Ollama, GitHub and other platforms. IBM also released Granite Speech 5.0 Turbo CTC models with 470 million parameters, claiming they can transcribe three hours of audio in one second.

Sources

  1. arstechnica.comIBM's new Granite 4.2 models ride the wave of interest in local LLMs
  2. the-decoder.comIBM drops open-weight Granite 4.2 family with built-in agentic capabilities under Apache 2.0

Related coverage

Continue reading