Granite 4.2 stays dense while rivals go hybrid

IBM released Granite 4.2 on Tuesday in 3 billion, 8 billion and 30 billion parameter sizes, all of them dense, decoder-only reasoning models pre-trained from scratch on 15 trillion tokens. The Apache 2.0 release is text-only and arrived alongside two new Granite Speech recognition models, according to The New Stack.

At a glance

• Thinking and non-thinking modes are joined by a low-effort setting that spends a small number of reasoning tokens on easy questions, making the reasoning behaviour optional rather than always on.
• Pre-training ran in five phases and included 1 trillion tokens of synthetic code from IBM's CodeAlchemy pipeline; the long-context phase takes the family to 512,000 tokens, with 128K native in the released configuration.
• The 8B model often lands close to the 30B on benchmarks and runs on virtually any modern Mac and some lower-end Nvidia RTX GPUs, according to The New Stack.

The architecture choice reads as a deliberate bet against the field's drift toward hybrid stacks: a plain attention Transformer is easier to fine-tune and deploy, and IBM has now walked that path twice in a row. The reasoning modes matter more than the benchmark ranking here, since IBM spent the 4.1 cycle arguing that enterprises could skip reasoning entirely. Adding it back, but optionally, likely reflects how much agentic work now depends on multi-step planning.

Granite 4.1 returned the family to all-attention dense Transformers

Much of the field has moved from all-attention Transformers toward hybrid Mamba/attention designs, among them Nvidia's Nemotron 3 family. IBM went down that road with Granite 4.0, a generation that shipped conventional dense models, dense hybrids and hybrid mixture-of-experts variants side by side.

With Granite 4.1 the main family went back to an all-attention, dense Transformer architecture. IBM said at the time that the newer models outperformed the previous generation while using a simpler and therefore more flexible architecture for fine-tuning on downstream tasks. The 4.2 line keeps that design across all three sizes.

IBM calls 4.2 a reasoning-focused release. When 4.1 launched, the company argued that reasoning models were not efficient enough and that cheaper non-reasoning models with similar benchmark results made sense for enterprise tasks such as instruction following and tool calling.

The 8B and 30B models got an extra agentic reinforcement learning stage

All three models share the same pre-training pipeline: 15 trillion tokens across five phases, including a long-context phase, and 1 trillion tokens of synthetic code generated by IBM's CodeAlchemy pipeline. The released configuration natively supports a 128K context window, with training taking the family to 512,000 tokens.

The 8B and 30B models went through an additional agentic reinforcement learning step that trains them to call tools, edit and run code, work in a terminal and search the web. Combined with RLHF alignment, IBM says the approach produces models better equipped for complex, multi-step agentic work.

The 3B model supports tool calling too, though it did not go through the extra agentic stage, and The New Stack cautions against expecting much of it. IBM describes 4.2 as aimed at agents that plan, call applications and execute multi-step tasks reliably.

Granite 4.2 ships text-only while rivals in its size class are multimodal

Granite 4.2 is text-only. Comparable models in the same size class, including Qwen-3.8 27B, Muse Glimmer 30B and Google's Gemma 4 31B, are multimodal. IBM does have a separate vision model in Granite Vision 4.1 4B, released under the previous generation.

On benchmarks the family breaks no new ground, according to The New Stack, which reports that Qwen 3.8 27B beats the Granite models across the board and that Granite's coding results are inconsistent, despite the trillion tokens of synthetic code in pre-training.

IBM's argument for Granite rests on performance in high-throughput agentic tasks rather than leaderboard position, and the announcement says an AI system must plan, call applications and execute complex tasks reliably while staying light enough to use without breaking the bank.

This release extends the Granite family with a clear goal: helping enterprises build agents that can reason, act, and adapt during real-life workflows.

Whether a 4.2 vision model follows

IBM has not said whether Granite Vision will get a 4.2 update; the current vision model remains Granite Vision 4.1 4B, and the 4.2 language models are text-only. Also unaddressed is when the released configuration might expose the full 512,000-token window rather than the 128K it natively supports, and no timeline has been given.