IBM Granite 4.1 8B: Openness as a Certified Property

What Granite 4.1 8B is

Granite-4.1-8B is IBM's 8 billion parameter long-context instruct model, released on 29 April 2026 under Apache 2.0. It is fine-tuned from Granite-4.1-8B-Base using a combination of open-source instruction datasets under permissive licences and IBM's own internally collected synthetic data, then put through supervised fine-tuning and reinforcement-learning alignment. IBM name the improvements that post-training targeted: tool calling, instruction following and chat. It carries a 128K token context window and runs at roughly $0.05 per million input tokens and $0.10 output through third-party providers, which makes it one of the cheapest models on this site.

The openness claim is externally certified

Most open-weight releases are open in the sense that you can download the weights. Granite 4.1 carries a Model Openness Framework Class 3 qualification, an external assessment of how much of a model's production was actually published — not just weights but training methodology, datasets, documentation and licensing. IBM publish the supporting material: a technical blog, a GitHub repository for the language models, and product documentation, all linked from the card.

Languages and intended use

IBM state support for twelve languages: English, German, Spanish, French, Japanese, Portuguese, Arabic, Czech, Italian, Korean, Dutch and Chinese. They also state plainly that users may fine-tune the model for languages beyond that list. The intended use is described as following general instructions and serving as a foundation for further work, which is consistent with a small model released alongside its own base checkpoint.

LMArena

On LMArena's text leaderboard Granite 4.1 8B has 1304 Elo (rank 286); by category, 1353 Elo in Coding (rank 271), 1317 Elo in Math (rank 236), 1339 Elo in Expert (rank 234) and 1260 Elo in Creative Writing (rank 291).

IBM’s benchmark table

IBM's card compares the Granite 4.1 dense models at 3B, 8B and 30B. For the 8B it reports MMLU 73.84 (5-shot), MMLU-Pro 55.99, BBH 80.51, AGI Eval 72.43, GPQA 41.96 and SimpleQA 4.82, with AlpacaEval 2.0 at 50.08 and an IFEval average of 87.06. The 3B records 67.02 on MMLU and 82.30 on IFEval; the 30B 80.16 and 89.65. These are IBM's own figures.

Tool calling and chat format

Tools are defined with OpenAI's function definition schema. The chat template lists available tools inside <tools> tags in the system turn, and the model returns each call as a JSON object with the function name and arguments inside <tool_call> tags; if a requested tool does not exist, the template instructs it to tell the user it cannot fulfil the request. Roles are delimited with <|start_of_role|> and <|end_of_role|> tokens. IBM describe the model as suited to AI assistants across domains, including business applications, and to LLM agents equipped with tool use. It is part of the Granite 4.1 language model collection on Hugging Face, with documentation on IBM's Granite docs site, a technical blog on Hugging Face and a GitHub repository, ibm-granite/granite-4.1-language-models. IBM's card shows the model answering a single-fact question in its chat format, and running through Hugging Face Transformers with a standard chat template.

Official Resources