01
What IBM released
IBM released Granite 4.2 on August 25, 2026 as a family of dense reasoning language models in 3B, 8B and 30B sizes. IBM positions the family for agentic and enterprise workflows where reasoning, tool use, coding and deployability matter.
IBM's current research post, Hugging Face repositories and GitHub organization are the strongest live references. IBM's older Granite documentation page now warns that it is no longer being updated, so use it as supporting context rather than the sole source for future changes.
02
3B, 8B and 30B model sizes
- Granite 4.2 3B: compact option for constrained and edge-oriented deployments.
- Granite 4.2 8B: balanced general-purpose reasoning model.
- Granite 4.2 30B: flagship model for deeper reasoning and specialized tasks.
The multiple sizes let teams reserve larger models for harder tasks while using smaller variants where latency, throughput or memory matters more.
03
Full thinking, low-effort and non-thinking modes
Granite 4.2 supports native reasoning and exposes flexible thinking behavior. IBM documents full thinking, non-thinking and low-effort modes so applications can trade reasoning depth against latency.
That flexibility matters in production because simple requests do not always justify the same token or latency budget as complex planning and coding tasks.
04
Tool calling and agentic training
IBM says all three models support reasoning-augmented tool calling. The 8B and 30B variants also received additional agentic reinforcement-learning training focused on software engineering, terminal interaction and search-driven workflows.
IBM documents vLLM serving through an OpenAI-compatible endpoint and model-specific reasoning/tool parsing. That makes Granite 4.2 relevant to teams building coding agents, RAG systems and internal automation.
05
Context window
IBM's family documentation states that all Granite 4.2 models natively support a 128K context window, with a long-context extension to 512K on the 30B model.
06
Open weights and Apache 2.0
IBM publishes Granite 4.2 checkpoints through its Hugging Face organization and releases the language models under Apache 2.0. IBM describes the family as open and suitable for research and commercial deployment subject to the license.
IBM also publishes quantized variants, including official GGUF repositories, which makes local and desktop experimentation more practical than releases that depend only on full-precision server checkpoints.
07
Can Granite 4.2 run locally?
Yes. IBM documents Transformers and vLLM usage, the Granite project describes additional serving support, and IBM-maintained GGUF repositories provide llama.cpp-compatible quantized paths.
For model selection, runtimes and realistic memory tradeoffs, read the Granite 4.2 local-running guide.
08
How to read IBM's benchmarks
IBM reports Granite 4.2 results across reasoning, coding, instruction following, tool calling and agentic benchmarks including AIME25, LiveCodeBench, BFCL and software-engineering evaluations.
09
Who should consider Granite 4.2?
The strongest early fit is for developers who want downloadable reasoning models for coding tools, tool-calling agents, RAG systems and enterprise workflows where deployability and Apache 2.0 licensing matter.
The 3B model is attractive for constrained deployments, 8B is the most practical general-purpose starting point, and 30B is the higher-capability option when the infrastructure budget supports it.
Sources
Primary and supporting sources
Facts were rechecked against the linked sources immediately before publication. Pricing, product availability and rollout status can change.