IBM Ships Granite 4.2: Three Dense Reasoning LLMs That Can Actually Drive a Terminal

What happened

IBM's Granite Team released Granite 4.2 on 25 August 2026: three dense, decoder-only reasoning LLMs at 3B, 8B, and 30B parameters, trained from scratch on roughly 15 trillion tokens. All three carry a thinking/non-thinking toggle, native tool calling in OpenAI function-calling format, and the Apache 2.0 licence.

Context

Granite 4.1 was primarily an instruction-following assistant family. 4.2 is the reasoning step: explicit chain-of-thought is the headline addition, and the five-phase pre-training recipe (foundational training, data annealing, long-context extension to 512K tokens) closely follows the 4.1 pipeline. The SFT corpus spans about 7.2 million samples, roughly 100 billion tokens, split 31.6 percent agentic and 68.4 percent non-agentic. The agentic side is dominated by software-engineering tasks: 69 percent SWE, 12.1 percent tool calling, 8.0 percent terminal use. The shift from following instructions to reasoning and acting is the whole story.

How it works

Architecture: GQA (40 query / 8 KV heads), RoPE θ = 10,000,000, SwiGLU MLPs, RMSNorm (ε = 1e-5), untied embeddings, bfloat16. Pre-training runs five phases with distinct data mixtures and learning-rate schedules, shifting from broad web-scale data toward curated sources; phase 5 extends context to 512K tokens.

SFT quality control uses GPT-OSS-120B and Gemma 4 as LLM-based judges to drop hallucinated or invalid samples; deduplication is SHA-256 over the tools-and-messages fields. Agentic samples were generated through scaffolds including OpenHands, SWE-agent, Terminus-2, OpenSeeker, and Goose.

The 8B and 30B receive an additional agentic RL block: tool calling, code execution, terminal driving, web search in sandboxed environments. The 3B does not. All models ship a low-effort thinking mode alongside full thinking.

Serving: OpenAI-compatible via vLLM or SGLang. Tool calls use the standard OpenAI function-calling schema.

Our read

The most useful fact is buried in the SFT section: GPT-OSS-120B and Gemma 4 serve as LLM-based judges to filter IBM's own training data. They are not grading their homework with their own models, and that is a quality signal.

The 8B is the size that matters for a small studio. It is the first tier that gets the agentic RL block: tool calling, code execution, terminal driving, web search in a sandbox. That maps onto running a ComfyUI workflow, iterating a negative prompt, batch-rendering a shot list without a human clicking next. At bfloat16, roughly 16–20 GB of VRAM, it fits a 24 GB card. The 3B, the only size realistic on a consumer GPU, lacks that stage. It has the native tool-calling interface but has not been RL-trained to chain calls. The gap between calling a tool and running a pipeline is the gap between those two sizes.

What the brief does not contain matters. No benchmarks, no throughput numbers, and the source is truncated before the full RL-stage description. You cannot tell from this post whether the 30B beats the 8B on agentic tasks or whether 512K context holds quality at the far end. IBM is publishing the recipe, not the taste test.

The OpenAI-compatible serving layer is the quiet win: no custom adapter, no proprietary SDK. If you already run vLLM or SGLang, Granite 4.2 is a drop-in.

What this changes

Pull the 8B, point your orchestration layer at it, and swap your prompt-based assistant for a model RL-trained to call tools and run code in a sandbox. Your existing tool-schema code does not change. Use low-effort thinking for tagging and captioning; reserve full thinking for script architecture.

On a single consumer GPU, the 3B is your ceiling: prompt engineering and simple JSON generation, no autonomous tool chaining. Multi-GPU or cloud: the 30B is the strongest option, and 512K context removes truncation for long storyboards.

No video-generation capability is added. This is an orchestration model, not a diffusion model. Your ComfyUI stack stays as is.

License

All three Granite 4.2 models are released under Apache 2.0. Commercial use is permitted without attribution, revenue-sharing, or model-size ceiling. You can embed the 8B in a client-delivered pipeline, serve it behind a product, or fine-tune and redistribute without licence friction.

Key takeaways

• Granite 4.2 (3B / 8B / 30B) is a dense, decoder-only reasoning family trained from scratch on ~15T tokens, with 512K context and native tool calling in OpenAI function-calling format.
• The 8B is the first size that receives agentic RL and fits on a single 24 GB card at bfloat16; the 3B lacks that stage.
• SFT quality control uses GPT-OSS-120B and Gemma 4 as external judges; the agentic corpus is 69% SWE, 12.1% tool calling, 8.0% terminal use.
• Apache 2.0: commercial use, fine-tuning, and redistribution are permitted without additional terms.
• No benchmark scores or throughput numbers are available in the published brief; the source is truncated before the full RL description.

Sources

1. Granite 4.2 LLMs: How They're Built — tier 1