Granite 4.2
IBM's open reasoning language model family, released August 2026
Granite 4.2 is a generation of open large language models from IBM's Granite team, released on 25 August 2026. The 30B flagship is a dense decoder model with built-in chain-of-thought, aimed at reasoning, code, tool use and multilingual dialogue, and shipped under Apache 2.0.
- Released 25 August 2026 under Apache 2.0
- 30B flagship is a dense decoder-only transformer
- Context of 128K tokens, extendable to 512K
- Three thinking modes: full, non-thinking, low-effort
- Trained on an NVIDIA GB200 NVL72 cluster at CoreWeave
What it is
Granite 4.2 is a family of open large language models from IBM, meant for business and research work. The 30B member is described as the top reasoning model of the group. It is a dense decoder-only transformer, so every parameter is used on each step, and it carries about 30 billion of them. IBM positions the line as covering small edge devices up to large reasoning systems. The 30B model targets reasoning, code writing, tool calls, agent-style workflows and conversation in several languages. It came out on 25 August 2026 under the Apache 2.0 licence, which allows commercial and academic use without restriction.
How it came about
The Granite line is IBM's family of open large language models. Granite 4.2 is the next generation after 4.1, and the 30B model was post-trained from Granite-4.1-30B-Base rather than trained from scratch. IBM says the 4.2 release adds native thinking, meaning the model works through steps before answering. The 30B model appeared on 25 August 2026 alongside 3B and 8B siblings in the Granite 4.2 collection. Hyperparameters were tuned first and then applied across all three sizes.
Inside the model
The model is a decoder-only dense transformer. It uses grouped query attention with 32 attention heads and 8 key-value heads, rotary position embeddings with theta 10,000,000, a SwiGLU feed-forward layer with hidden size 32768, RMSNorm at epsilon 1e-5, and separate input and output embeddings. Post-training ran in three stages: a multilingual pre-training base, supervised fine-tuning on instruction, reasoning and chain-of-thought data, then reinforcement learning with Group Relative Policy Optimization across maths, code, science, tool use and chat environments. Rewards were verifiable where possible and model-scored otherwise. Training used an NVIDIA GB200 NVL72 cluster at CoreWeave. A later RLHF pass tuned helpfulness, conversation quality and safety.
What it can do
IBM says built-in chain-of-thought raises performance on maths, code and multi-step problems. Three thinking settings are offered: full reasoning, no reasoning, and a low-effort mode for simple queries. The model can reason about which tool to call before emitting a function call. Context runs to 128K tokens natively and 512K after long-context extension. Tested languages include English, German, Spanish, French, Japanese, Portuguese, Arabic, Czech, Italian, Korean, Dutch and Chinese; IBM notes other languages may work but were not fully tested. Evaluations use a framework built on NVIDIA's NeMo Evaluator SDK.
How to use it
Weights are published under Apache 2.0, so anyone can use them commercially or academically. IBM recommends serving with vLLM, using a custom reasoning parser called granite_thinking_parser and the qwen3_coder tool-call parser. The model also runs with the built-in nemotron_v3 parser. IBM says native parser support is planned for vLLM and SGLang. Suggested generation settings are temperature 1.0 and top_p 0.95 for every task.
Safety and testing
A preference-alignment stage after the reward-driven training phases tunes helpfulness, conversational quality and safety. IBM describes the Granite family as built to be safe as well as versatile and efficient. No further safety evaluations or incidents are described here.
Why it matters
Granite 4.2 brings native step-by-step reasoning into IBM's open model line, which the company says lifts results on complex maths, coding, multi-step logic and agentic tool use. The models sit in a family that spans compact edge-deployable sizes up to large reasoning systems, all released under Apache 2.0 for unrestricted use.
Questions people ask about Granite 4.2
What is Granite 4.2?
It is IBM's newest generation of open large language models, released on 25 August 2026. The 30B flagship is a dense decoder model with built-in chain-of-thought, aimed at reasoning, code, tool calls and multilingual dialogue, and published under Apache 2.0.
How large is the model and how much context does it handle?
The flagship carries about 30 billion parameters. It accepts 128K tokens of context natively, and IBM says that can be extended to 512K for long documents, extended conversations and agent-style work.
Can I use it commercially?
Yes. The weights come under the Apache 2.0 licence, which IBM describes as allowing unrestricted commercial and academic use. The models are downloadable rather than API-only.
What is special about its reasoning?
Reasoning is built in rather than bolted on. The model can work through steps inside a thinking block before answering, and users can pick full thinking, no thinking, or a low-effort setting to trade depth against speed on each query.
Which languages does it support?
Testing covered English, German, Spanish, French, Japanese, Portuguese, Arabic, Czech, Italian, Korean, Dutch and Chinese. IBM cautions that other languages may work but were not fully tested.
Good reading and sources
There is no English Wikipedia article on Granite 4.2 yet, so the facts on this page come from the sources listed above: huggingface.co/ibm-granite/granite-4.2-30b. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.