What it is
Granite 4.2 is IBM's open model family, released on 25 August 2026. It comes in three dense sizes, 3B, 8B and 30B, all text in and text out. Each model can think step by step, answer directly, or use a low-effort mode that reasons briefly.
All sizes were trained with reinforcement learning on maths, science, coding and tool calling; the 8B and 30B also got agent training on software engineering, terminal work and web search. IBM added 1 trillion tokens of synthetic code. The models handle 128K tokens natively, and the 8B and 30B cards list an extension to 512K. They support 12 languages, including English, German, Spanish, French, Japanese, Arabic, Korean and Chinese.
The weights are under Apache 2.0 on Hugging Face, Ollama and LM Studio, and the models are hosted on watsonx, OpenRouter, Replicate, DeepInfra, CoreWeave and others.
Our review
Granite 4.2 is a sound, unglamorous choice: three dense models under a clean licence that reason well for their size and land mid-table when someone else does the scoring. Artificial Analysis gives the 30B 13 on its Intelligence Index, 29th of 141 open models between 4B and 40B parameters, where the median is 8. That is one point above OpenAI's gpt-oss-120b, a model four times its size.
IBM's card has the rest, all self-reported on a harness built on NVIDIA's NeMo Evaluator, with no rival models in the table. The 30B reaches 57 on SWE-bench Verified, 29.24 on Terminal-Bench 2.1 and 89.17 on AIME 2025; the 8B scores 47.67 on SWE-bench Verified. One number undercuts the agent story: on the BFCL v4 tool-calling test the 8B scores 52.39, level with the 3B's 52.41, despite the extra agent training. Long context holds up on the 30B (81.38 on RULER at 128K) and fades on the 3B (55.30).
The models are easy to run. Ollama's 4-bit builds are 2.2GB, 5.3GB and 18GB, so the 30B fits a 24GB card and the 8B a laptop. vLLM needs version 0.20 or later, a custom reasoning parser from the repository and the qwen3_coder tool parser. The 512K extension is harder to find in practice: the card's own serving examples stop at 131,072 tokens, and every Ollama tag lists 128K.
Hosted, it is cheap but chatty. Artificial Analysis tracks one API provider, IBM, at $0.16 per million input tokens and $0.65 per million output, at 67 tokens a second, below its class median. The 30B used 100M output tokens to complete the index against a median of 82M, so the non-thinking and low-effort switches are the main cost controls.
It is text-only, which leaves images and audio to Gemma 4. For a company that already buys watsonx, though, the same weights arrive inside an existing contract, and that is a shorter procurement path than any leaderboard position.
Where it earns its place
A platform team at an insurer running agents on premises
Run a tool-calling claims assistant on one 48GB GPU
The 30B at 8-bit is a 31GB build, leaving room for context on a 48GB card, and scores 61.39 on BFCL v4. Apache 2.0 removes licence review, and vLLM's tool parser handles function calls.
A developer with a 16GB laptop
Use a local coding assistant offline
The 8B at 4-bit is a 5.3GB Ollama download; 47.67 on SWE-bench Verified by IBM's count is useful for small fixes but well short of hosted coding models. The 3B, at 2.2GB, suits quick questions.
A product manager pricing an AI triage feature on watsonx
Classify and summarise 1M support requests a month
At 1,500 input and 800 output tokens a request, IBM's $0.16/$0.65 rates come to about $760 a month. Switching thinking off for routine tickets cuts the output share, which is most of the bill.
Before you commit
- Setup
- ollama run granite4.2 pulls the 8B; vLLM 0.20+ needs the repository's granite_thinking_parser file and --tool-call-parser qwen3_coder; SGLang 0.5.18+ also works.
- Integrations
- Hugging Face, Ollama, LM Studio, watsonx, OpenRouter, Replicate, DeepInfra and CoreWeave; thinking modes set through chat_template_kwargs.
- Data and privacy
- Apache 2.0, open for commercial and research use with no field-of-use limits; self-hosted prompts stay on your hardware.
- Cost at scale
- IBM's API charges $0.16 per million input and $0.65 per million output tokens, with a 75% discount on cache hits. Self-hosted, the 30B needs about 18GB at 4-bit or 31GB at 8-bit.
- Limits
- 128K context in practice; the 512K extension is not in the reference serving setup. Text only. 67 tokens a second on IBM's API, and verbose when thinking.
What's good
- Apache 2.0 with no conditions
- Dense 3B, 8B and 30B; 30B fits a 24GB GPU at 4-bit
- Thinking, non-thinking and low-effort modes in one model
- 30B scores 13 on Artificial Analysis, above gpt-oss-120b
What's not
- Text-only
- Card benchmarks are IBM's own, with no rival comparisons
- 8B no better than 3B on BFCL tool calling
- Verbose in thinking mode, which raises output bills
Best for
Companies and developers who want small, permissively licensed reasoning models for local or on-premises agents.
Alternatives
Granite 4.2: common questions
- Is Granite 4.2 worth it?
- Granite 4.2 is the open family a risk committee will approve fastest, with reasoning that is good for its size and nowhere near the frontier. We rate it 4.0 out of 5.
- How much does Granite 4.2 cost?
- Free weights under Apache 2.0; watsonx and third-party hosting billed separately.
- What are the downsides of Granite 4.2?
- Text-only; card benchmarks are IBM's own, with no rival comparisons; 8B no better than 3B on BFCL tool calling; verbose in thinking mode, which raises output bills.
- Who is Granite 4.2 best for?
- Companies and developers who want small, permissively licensed reasoning models for local or on-premises agents.
- How much does Granite 4.2 cost for a team or in production?
- IBM's API charges $0.16 per million input and $0.65 per million output tokens, with a 75% discount on cache hits. Self-hosted, the 30B needs about 18GB at 4-bit or 31GB at 8-bit.
- Is Granite 4.2 safe for company data?
- Apache 2.0, open for commercial and research use with no field-of-use limits; self-hosted prompts stay on your hardware.
- What are the best Granite 4.2 alternatives?
- The closest alternatives we have reviewed are Gemma 4, gpt-oss-20b, Devstral Small 2 and Falcon-H1R-7B.
Answers checked Oct 22, 2026.
Make Granite 4.2? Show this rating on your site
