Is IBM Granite free? Yes, for almost everyone. The current IBM Granite language models (Granite 4.2 in 3B, 8B and 30B sizes, plus the Granite 4.0 Nano models) are released under the Apache 2.0 license, so you can download them, run them on your own hardware, fine tune them and ship them in a paid product without paying IBM. You pay only if you use IBM's hosted service, watsonx.ai, beyond its free monthly allowance, or if you pick one of the few Granite checkpoints that carry a different license.
This guide explains exactly what "free" covers, where the exceptions are, and when paying IBM gets you something the free download does not. Every license, size and price below comes from IBM's own pages, IBM's Granite organization on Hugging Face and the Ollama library, all checked on November 12, 2026. We did not run Granite for this guide.
What IBM Granite is
Granite is IBM's family of open models for business work. The Granite home page groups it into six lines: Language, Speech, Vision, Guardian (guardrail models that flag harmful prompts and outputs), Embedding and Time Series. Most people asking whether Granite is free mean the language models, the ones you chat with or wire into an agent.
The newest language release is Granite 4.2, which IBM Research announced on August 25, 2026. According to the Granite 4.2 documentation, it comes in three dense sizes:
| Model | Parameters | Context | Ollama download (default tag) | Aimed at |
|---|---|---|---|---|
granite-4.2-3b | 3B | 128K | 2.2 GB | Edge devices and small machines |
granite-4.2-8b | 8B | 128K | 5.3 GB | General business use |
granite-4.2-30b | 30B | 128K, extendable to 512K | 18 GB | Harder reasoning and coding |
All three have a built in thinking mode that you can switch on, switch off or set to a low effort setting per request, and they support tool calling. The model cards list 12 tested languages: English, German, Spanish, French, Japanese, Portuguese, Arabic, Czech, Italian, Korean, Dutch and Chinese. The download sizes are the default granite4.2 tags on the Ollama library, which are 4 bit quantized files.
What the Apache 2.0 license lets you do
Apache 2.0 is one of the most permissive licenses in software. For Granite weights it means:
- No fee. There is no charge to download, use or redistribute the weights.
- Commercial use is allowed. You can build a paid app or service on Granite. The Granite 4.2 8B model card states the license enables "unrestricted commercial and academic use."
- No user count limit. Unlike the Llama community licenses, Apache 2.0 has no clause about monthly active users.
- No gated download. The Granite 4.2 repositories on Hugging Face open without a request form or an account.
- Fine tuning and redistribution are allowed. You can release your own fine tuned version, as long as you keep the license text and notices with it.
The license also has limits that matter in practice. It gives you the weights "as is," with no warranty. It does not grant rights to IBM's trademarks, so you should not name your product in a way that suggests IBM made it. And on the Granite home page, IBM adds a rule that Granite users may not use the models, or let others use them, to generate output that infringes anyone else's copyright. That rule is on a web page rather than in the license file, but it tells you how IBM expects the models to be used.
If you want the bigger picture on how Apache 2.0 weights differ from fully open models with public training data, our explainer on open weight vs open source covers it.
Which Granite models are not free for commercial use
Most of IBM's Granite organization on Hugging Face is Apache 2.0, but not all of it. When we listed IBM's newest 40 repositories on November 12, 2026, two used other licenses:
| Model | License | What it means |
|---|---|---|
granite-speech-5.0-470m-turboctc-nc | CC BY-NC-SA 4.0 | Research and non commercial use only. The card points business users to the Apache 2.0 granite-speech-5.0-470m-turboctc instead. |
granite-timeseries-patchtst-fm-r2 | OpenMDW 1.0 | A different open license for model weights. Read it before you ship. |
The rule is simple: check the license: field on each model card before you build on it. The language models you are most likely to use (Granite 4.2, 4.1 and the 4.0 Nano and 4.0 families) were all Apache 2.0 when we checked.
How to use IBM Granite for free
You have three free routes, and none of them needs an IBM account.
Run it locally with Ollama
This is the quickest way. Install Ollama, then pull a size that fits your machine:
ollama run granite4.2:8bSwap in granite4.2:3b for a laptop with little memory or granite4.2:30b if you have a 24 GB graphics card or a Mac with plenty of unified memory. The Ollama tags page lists smaller quantized versions too, down to a 1.5 GB granite4.2:3b-q2_K file. Our guide to how much VRAM you need for an LLM explains how to match a model size to your hardware.
Download the GGUF or MLX files
IBM publishes its own GGUF builds for llama.cpp and LM Studio. The granite-4.2-8b-GGUF repository runs from a 3.41 GB Q2_K file to a 17.59 GB bf16 file, with Q4_K_M at 5.35 GB. IBM also uploaded MLX versions for Apple silicon in bf16, 8 bit and 4 bit. If the quant names are new to you, see GGUF quantization explained, and for llama.cpp setup see how to use llama.cpp.
Serve it with Transformers or vLLM
For your own API, load the full weights with Hugging Face Transformers or serve them with vLLM. The Granite 4.2 docs say the models are tuned for vLLM 0.20 or newer, with the granite_thinking_parser (shipped in each model repository) for reasoning output and the qwen3_coder parser for tool calls. vLLM then exposes an OpenAI compatible endpoint on your own server. IBM recommends a temperature of 1.0 and top p of 0.95 for every task.
Because the weights are Apache 2.0, you can also fine tune Granite on your own data with tools such as Unsloth. Our guide to fine tuning an LLM locally walks through the process.
IBM Granite 4 Nano: the smallest free models
If you searched for "IBM Granite 4 nano," these are IBM's smallest language models. The Granite 4.0 H 1B model card gives a release date of October 28, 2025, an Apache 2.0 license and the same 12 supported languages. The Nano collection has 1B and 350M sizes, each in a hybrid ("H") version and a standard transformer version.
On the Ollama granite4 tags page, granite4:350m-h is 366 MB with a 1M token context window, granite4:1b-h is 1.6 GB, also with a 1M window, and the standard granite4:350m is 708 MB with a 32K window. Models this small suit phones, single board computers and quick classification or extraction jobs rather than open ended chat. For other tiny options, compare our SmolLM vs Gemma guide.
When IBM Granite costs money: watsonx.ai
IBM also hosts Granite on its cloud platform, watsonx.ai, and that is a paid service with a free allowance. The watsonx.ai pricing page lists a Free plan with up to 300,000 foundation model tokens per month, up to 20 compute usage hours of machine learning tools and up to 100 text extraction documents. Above that, you move to pay as you go (the Essentials plan, listed from USD 0 a month plus usage) or a higher tier listed from USD 1,110 a month. Per token charges use a resource unit of 1,000 tokens, counting input and output together.
Prices vary by model. IBM's supported models documentation lists granite-4-h-small at USD 0.0000636 per 1,000 input tokens and USD 0.000265 per 1,000 output tokens, which works out to about USD 0.064 and USD 0.265 per million. On November 12, 2026, the hosted catalog listed Granite 4.1 models (3B, 8B and 30B) and the Granite 4.0 H models, but not Granite 4.2, which you could still only download.
What paying IBM gets you: IP indemnity
The main reason a company might pay for Granite on watsonx.ai rather than run the free download is legal cover. On the Granite page, IBM says it gives its own models the same standard contractual intellectual property indemnification it offers for its hardware and software, that it does not ask customers to indemnify IBM in return, and that it does not cap that liability.
That protection depends on where you get the model. IBM's watsonx.ai documentation says IBM models you obtain from an outside repository such as Hugging Face are not indemnified by IBM, and that only IBM models you access from watsonx.ai are, with the terms set out in the IBM Client Relationship Agreement. So the same Granite weights come with IBM's contractual cover when you call them through watsonx.ai, and with only the Apache 2.0 terms (no warranty, no indemnity) when you download them from Hugging Face or Ollama.
For a hobby project or an internal tool, the free download is usually enough. For a regulated business that wants a vendor to stand behind the model, the paid route is what you are buying. Check the current contract terms with IBM before relying on them, because they sit in your agreement, not in the license.
Free vs paid Granite at a glance
| Item | Free download (Hugging Face, Ollama, GGUF, MLX) | watsonx.ai (IBM hosted) |
|---|---|---|
| Cost | None, you pay for your own hardware | Free plan up to 300,000 tokens a month, then per token or hourly |
| License | Apache 2.0 for the main language models | IBM service terms |
| Newest language models | Granite 4.2 (3B, 8B, 30B) | Granite 4.1 and 4.0 H when we checked |
| IP indemnity from IBM | No | Yes, uncapped per IBM, under your agreement |
| Data leaves your machine | No | Yes, it runs in IBM's cloud |
| Account needed | No | IBM Cloud account |
Is IBM Granite good enough to use?
Free only matters if the model does the job. IBM publishes its own scores on the Granite page: on MMLU-Pro (5 shot), Granite 4.2 scores 67.84 for 3B, 74.04 for 8B and 77.60 for 30B, and on SWE-Bench Verified the 8B scores 47.67 and the 30B 57.00. These are IBM's figures, not independent tests, and we have not checked them. For independent rankings of open weight models, see our open weight model leaderboard, and for other free families worth trying, see open source LLM models to run locally.
Our pick: start with granite4.2:8b in Ollama if you have 8 GB or more of free memory, and drop to the 3B or a Nano model for small devices. Pay for watsonx.ai only when you need IBM's indemnity or a managed endpoint.
FAQ
Is IBM Granite free for commercial use?
Yes. The Granite 4.2, 4.1 and 4.0 language models are Apache 2.0, which allows commercial use, modification and redistribution with no fee and no user limit. Keep the license text with any copy you share, and check each model card, because a few checkpoints such as granite-speech-5.0-470m-turboctc-nc are non commercial only.
Is IBM Granite open source?
IBM calls Granite open source and releases the weights under Apache 2.0, an OSI approved license. Like most open weight models, the full training data is not something you can download and rebuild from, so some people prefer the term open weight. See our open weight vs open source explainer for the distinction.
Do I need an IBM account to use Granite?
No. The Hugging Face repositories for Granite 4.2 are not gated, and ollama run granite4.2:8b needs no sign in. You need an IBM Cloud account only for watsonx.ai.
How much does IBM Granite cost on watsonx.ai?
watsonx.ai has a free plan with up to 300,000 foundation model tokens a month. Beyond that, IBM bills per 1,000 tokens. For example, IBM's documentation lists granite-4-h-small at about USD 0.064 per million input tokens and USD 0.265 per million output tokens. Check the pricing page for the model you plan to use.
Does IBM indemnify Granite if I download it from Hugging Face?
No. IBM's watsonx.ai documentation says Granite models obtained from an outside repository such as Hugging Face are not indemnified by IBM. Only models you access through watsonx.ai carry IBM's indemnity, under the IBM Client Relationship Agreement.
What is the smallest free IBM Granite model?
The Granite 4.0 Nano 350M models. In Ollama, granite4:350m-h is a 366 MB download with a 1M token context window, small enough for a phone or a Raspberry Pi class board.
