Glonce

IBM ships Granite 4.2 reasoning models in three sizes

modelsThe Glonce editorsHugging Face🔥75Glonce rating75Impact72Novelty68Relevance98Surprise50Players72Credibility88
IBM ships Granite 4.2 reasoning models in three sizes

IBM's Granite Team has published Granite 4.2, described as the company's first family of dense, decoder-only reasoning LLMs. It ships in three sizes, 3B, 8B and 30B, each pre-trained from scratch on about 15 trillion tokens and released under the Apache 2.0 license. Every model in the family has a thinking and non-thinking switch, a low-effort thinking mode that spends a short reasoning budget on easy questions, and native tool calling.

Architecturally, all three sizes share the same decoder-only dense transformer design: Grouped Query Attention with 40 attention heads and 8 KV heads, Rotary Position Embedding with theta set to 10,000,000, an SwiGLU feed-forward block, RMSNorm with epsilon 1e-5, separate (untied) input and output embeddings, and bfloat16 precision.

Pre-training uses a five-phase strategy on the roughly 15 trillion tokens. Phases 1 and 2 cover foundational pre-training, phases 3 and 4 are mid-training with progressively higher-quality data annealing, and phase 5 is long-context training that extends the context window to 512K tokens. IBM says the pre-training recipe closely follows the previous generation, Granite 4.1.

Supervised fine-tuning uses a mixture of about 7.2 million samples, roughly 100 billion tokens of which about 65 billion are trainable, split 31.6% agentic and 68.4% non-agentic. The agentic slice breaks down into software engineering (69%), tool calling (12.1%), terminal use (8.0%), math (3.5%), search (0.8%) and action (0.2%), built with agent scaffolds including OpenHands, OpenCode, Terminus-2, SWE-agent, OpenResearcher, MiniSWE, OpenSeeker, EnvScaler, Gemini CLI, Hermes, Codex and Goose. The non-agentic slice covers instruction following (18.8%), coding (18.8%), math (14.6%), multilingual (7.0%), science (5.4%), reasoning (3.0%) and safety (0.8%). Before training, samples were normalized into a consistent chat format, scored by GPT-OSS-120B and Gemma 4 acting as LLM judges (removing low-scoring, hallucinated or invalid-tool-call samples), and deduplicated using SHA-256 hashes over the combined tools and messages fields, both within and across data sources. The 30B model additionally gets a second SFT phase focused on agentic coding, upsampling agentic, SWE and coding data while keeping about 16% of the mixture as replay from the original corpus, fine-tuned for roughly one more epoch at a learning rate of 3.0e-6.

After SFT, IBM runs a multi-stage, multi-environment reinforcement learning pipeline using asynchronous GRPO (Group Relative Policy Optimization), where a pool of generation workers keeps sampling while the trainer periodically pulls a batch and pushes updated parameters back, bounded by a limit on how far workers can drift from the current policy and a truncated-importance-sampling term in the objective to control stale samples. The staged curriculum runs SFT into RLVR, then skill boosters, then an agentic block (software engineering, terminal use, web search), then RLHF, with each stage warm-starting from the previous stage's checkpoint. The agentic block runs only for the 8B and 30B models; the 3B model takes a shorter path of foundational RL and alignment without it. For the RLVR stage on the 30B model, each step pairs 256 prompts with 16 sampled responses apiece for a 4,096-example batch. The KL-penalty coefficient varies by stage: it is 0 for RLVR and a second SWE stage, where rewards are objective and verifiable, and 0.05 for RLHF and a code-skill booster, where the objective is preference, safety or a narrow skill graft.

On deployment, Granite 4.2 emits tool calls in the OpenAI function-calling format when served through an OpenAI-compatible endpoint such as vLLM, so it plugs into agentic harnesses without extra glue code; IBM also points to a SGLang cookbook with a ready-to-serve recipe.

Key facts

  • Granite 4.2 ships in three sizes, 3B, 8B and 30B, each pre-trained from scratch on about 15 trillion tokens, released under Apache 2.0.
  • A five-phase pre-training strategy extends the context window to 512K tokens in its final phase.
  • The SFT mixture totals about 7.2 million samples (roughly 100B tokens, about 65B trainable), split 31.6% agentic and 68.4% non-agentic data.
  • Only the 8B and 30B models go through an agentic reinforcement-learning block covering software engineering, terminal use and web search; the 3B model skips it.
  • The RL pipeline runs on asynchronous GRPO, with the RLVR stage on the 30B model pairing 256 prompts with 16 sampled responses each for a 4,096-example batch.

Why it matters

Granite 4.2 marks IBM's shift from instruction-following assistants to an explicitly reasoning-focused release. Every model in the family can produce a chain of thought before answering and can switch between thinking and non-thinking modes depending on how much deliberation a task needs, with a low-effort mode in between for easy questions. Paired with native tool calling in the OpenAI function-calling format, the models are built to plug directly into agentic harnesses without extra glue code, which signals where IBM expects enterprise LLM demand to move.

Who it affects

Developers building on open-weight models get three sizes to choose from, 3B, 8B and 30B, all under the Apache 2.0 license and servable through vLLM or SGLang. Teams that specifically need agentic behavior, calling tools, editing and running code, driving a terminal, or searching the web inside real environments, should look at the 8B and 30B versions, since only those two went through the agentic RL block; the 3B model took a shorter foundational-RL path without it.

How to use it

All three models expose a thinking and non-thinking switch plus native tool calling. Served through an OpenAI-compatible endpoint such as vLLM, Granite 4.2 emits tool calls in the OpenAI function-calling format, and IBM points to a SGLang cookbook with a ready-to-serve recipe. The models are released under Apache 2.0; the source text gives no release date, pricing or hosting details beyond that.

How solid is it

IBM documents the build in detail: a five-phase pre-training run on about 15 trillion tokens, an SFT mixture of roughly 7.2 million samples filtered by GPT-OSS-120B and Gemma 4 acting as LLM judges to remove low-scoring or hallucinated samples, and SHA-256-based deduplication across the full mixture. The RL side follows a defined, warm-started curriculum from verifiable-reward RL through skill boosters, an agentic block, and RLHF. IBM also reports benchmark scores for all three sizes, including AIME25 at 78.33, 86.67 and 89.17, SWE Bench Verified at 47.67 for the 8B and 57.00 for the 30B, and MMLU-Pro at 67.84, 74.04 and 77.60. What is missing from this account is outside validation: no head-to-head comparisons against other model families appear in the text.

Risks and caveats

This description comes from IBM's own blog post, so the training pipeline and its data-quality claims are the company's characterization of its own model, not an independent audit. The text does not state a release date, pricing or hosting terms, hardware or cluster details, or why GPT-OSS-120B and Gemma 4 specifically were chosen as judge models, so those points remain open.