Granite 4.2 LLMs: How They're Built

IBM has released Granite 4.2, its first family of dense, decoder-only language models built specifically for reasoning, in 3B, 8B and 30B parameter sizes. A technical walkthrough from the Granite Team says the models were trained from scratch on about 15 trillion tokens, support reasoning controls and native tool calls, and are available under the Apache 2.0 license, giving developers broad permission to use and modify them.

According to IBM, each Granite 4.2 model follows the same broad development sequence: five-phase pretraining, supervised fine-tuning and a multistage reinforcement-learning pipeline. The first four pretraining phases move from broad web-scale material toward more curated data. The fifth adds long-context training up to 512,000 tokens. IBM describes the architecture as a dense transformer using grouped-query attention, rotary position embeddings, SwiGLU feed-forward layers, RMSNorm and bfloat16 precision.

The supervised fine-tuning mixture contains about 7.2 million samples, representing roughly 100 billion tokens, of which IBM says about 65 billion were trainable. Agent-oriented material accounts for 31.6% of the mixture and covers software engineering, tool calling, terminal work, mathematics, search and computer actions. The remaining 68.4% covers instruction following, coding, mathematics, multilingual tasks, science, reasoning and safety. IBM says it standardized samples into an OpenAI Chat-compatible format, filtered them with model-based judges and heuristic checks, and removed duplicates using SHA-256 hashes.

The main post-training difference concerns agent behavior. IBM says the 8B and 30B models received an additional agentic reinforcement-learning stage in which they called tools, edited and ran code, operated terminals and searched the web inside real sandboxed environments. All three models support native tool calling, but the source does not say that the 3B model received this same sandbox-based reinforcement-learning block.

Agent Training Moves Into Sandboxes

The release expands IBM’s Granite line from instruction-following assistants into models designed for explicit reasoning. Each model can run in thinking or non-thinking mode, while a low-effort setting gives simpler prompts a smaller reasoning budget. That control may help developers balance response speed and computational cost against the amount of deliberation a task receives.

The licensing and serving choices also affect adoption. The Apache 2.0 release permits commercial use and modification, while support for OpenAI-style function calls can reduce integration work for existing applications. IBM says the models can be served with vLLM or SGLang and connected to agent harnesses without a custom tool-call translation layer. The practical value will depend on measured reliability, hardware needs and performance outside IBM’s training environments.

Five Stages Shape Granite 4.2

Earlier Granite releases focused mainly on instruction following. Granite 4.2 adds a training program aimed at producing reasoning traces and tool-using behavior while retaining ordinary answer generation. IBM says the three sizes share the same overall architectural family and training sequence, although layer counts, hidden dimensions and parameter totals vary by model.

The 3B version has 40 layers and a 2,560-dimensional embedding, while the 8B model uses 40 layers with a 4,096-dimensional embedding. The 30B model increases the depth to 64 layers and uses a much larger feed-forward hidden dimension. IBM’s table lists a 131,072-token sequence length for each architecture, while the pretraining description says the final phase extends training to 512,000 tokens.

IBM says the agentic fine-tuning material came from open datasets and synthetically generated environments using harnesses including OpenHands, SWE-agent, Terminus-2, OpenResearcher and Codex. Software engineering represents 69% of the agentic corpus, making it by far the largest category. IBM also says GPT-OSS-120B and Gemma 4 were used as automated judges to reject low-quality samples, fabricated information and invalid tool interactions.

“Granite 4.2 is our first family of dense, decoder-only reasoning LLMs, released in three sizes: 3B, 8B, and 30B.”

— IBM Granite Team

Benchmark and Architecture Gaps Persist

The supplied technical account does not provide enough benchmark results to independently compare Granite 4.2’s reasoning quality with competing open models. It also does not quantify tool-call error rates, sandbox task success, inference cost or the effect of the three reasoning modes. IBM’s capability descriptions remain vendor-reported claims pending broader testing.

Two specifications also need clarification. The architecture summary describes 40 attention heads, while the accompanying table lists 32 for the 8B and 30B models. The source also lists a 131,072-token sequence length alongside 512,000-token long-context training without explaining whether those figures refer to different training stages, default serving limits or supported inference configurations.

Independent Tests Will Measure Reliability

Developers can now examine the released weights, documentation and code, then test the models through vLLM, SGLang and compatible agent frameworks. The next evidence is likely to come from independent evaluations of reasoning accuracy, coding performance, long-context retrieval and tool-use reliability.

IBM may also need to publish clearer configuration details and fuller evaluation results, especially for the 512K context claim and the differences between the 3B model and its larger counterparts. Real deployments will show whether sandbox-trained agent behavior transfers reliably to unfamiliar tools, repositories and workflows.

Key Questions

What models are included in Granite 4.2?

The family contains 3B, 8B and 30B parameter models. All three use dense, decoder-only transformer designs and support thinking, low-effort thinking and non-thinking modes.

How much data was used to train Granite 4.2?

IBM says each model was pretrained from scratch on approximately 15 trillion tokens. Supervised fine-tuning then used about 7.2 million samples, totaling roughly 100 billion tokens.

Can every Granite 4.2 model call tools?

IBM says all three models support native tool calling and can emit OpenAI-compatible function calls. Only the 8B and 30B versions are described as receiving the separate agentic reinforcement-learning stage in live sandbox environments.

Does Granite 4.2 support a 512K context window?

IBM says the fifth pretraining phase extended the context window to 512,000 tokens. Its architecture table separately lists 131,072 tokens, and the supplied source does not reconcile the two figures.

What license applies to Granite 4.2?

IBM released the models under the Apache 2.0 license. That license generally permits use, distribution and modification, including in commercial applications, subject to its notice and licensing conditions.

Source: Hugging Face