IBM released the Granite 4.2 family of open-weight language models on Tuesday, arriving in 3 billion, 8 billion and 30 billion parameter sizes. The release is notable less for its benchmark scores than for the architectural road not taken: while much of the field has drifted toward hybrid Mamba and attention designs, IBM shipped dense, decoder-only reasoning models pre-trained from scratch.
That is a deliberate reversal. The Granite 4.0 generation included dense, dense-hybrid and hybrid mixture-of-experts variants, following the same direction as families like Nvidia's Nemotron 3. With Granite 4.1, IBM pulled the main line back to an all-attention dense transformer, arguing that a simpler architecture is more flexible to fine-tune for downstream work. Granite 4.2 stays on that path.
Reasoning arrives, and stays optional
IBM describes 4.2 as a reasoning-focused release, which marks a shift in the company's own position. When Granite 4.1 shipped, IBM argued that reasoning models were not yet efficient enough for enterprise use and that cheaper non-reasoning models made more sense for instruction following and tool calling. The team now treats reasoning as necessary, though it left the switch in the user's hands.
Every model carries a thinking and non-thinking toggle plus a low-effort mode that spends only a short reasoning budget on easy questions, along with native tool calling. The training pipeline runs roughly 15 trillion tokens across five phases, including a long-context stage that extends the family to a 512,000-token window, although the released configuration natively supports 128K.
What the bigger models get
All three sizes share the base pipeline, but the 8B and 30B models pass through an additional agentic reinforcement learning phase aimed at enterprise-style tasks: calling tools, editing and running code, working in a terminal and searching the web. Combined with reinforcement learning from human feedback, IBM says the result is a model better suited to multi-step agentic work. The 3B model supports tool calling as well, with correspondingly modest expectations.
Two further training choices shaped the coding side. The models saw 1 trillion tokens of synthetic code generated through IBM's CodeAlchemy pipeline, and a speculative decoding layer lets them emit text faster while serving more users, which cuts operating cost per request.
Honest about the benchmarks
Granite 4.2 does not top the charts, and IBM's framing does not pretend otherwise. Qwen 3.8 27B outperforms the Granite models broadly, particularly on coding, where the IBM family delivers inconsistent results. Granite is also text-only, unlike similarly sized multimodal rivals such as Qwen 3.8 27B, Muse Glimmer 30B and Google's Gemma 4 31B, though IBM maintains a separate Granite Vision line.
The more interesting detail is internal: the 8B model lands close to the 30B on many evaluations, and it runs comfortably on a modern Mac or a mid-range Nvidia RTX card. For high-throughput agentic workloads where a frontier model is overkill, that ratio is the actual argument. Every model ships under Apache 2.0 and is downloadable from Hugging Face, Ollama and GitHub.
Speech models tag along
IBM also released two speech recognition models the same day, named Granite Speech 5.0 Turbo CTC and a non-commercial variant trained on restricted-use data. At roughly 470 million parameters they are among the smallest models the company ships, and they drop the language model backbone entirely in favour of connectionist temporal classification.
The efficiency claim is striking. IBM's researchers measured a real-time factor near 12,600 on a single H200, against roughly 6,000 for the current speed leaders on the Hugging Face open transcription leaderboard, fast enough to process three hours of recorded audio in about a second.
Taken together, the release reads as a wager that enterprises will pay for predictable, licence-free models that fit on existing hardware rather than chase leaderboard positions, and IBM has separately begun working with Hirundo on machine unlearning to trim undesirable outputs without full retraining.
