Publication date: May 1, 2026
IBM has released the Granite 4.1 family of language models in three sizes: 3B, 8B, and 30B parameters, all available under the Apache 2.0 license. The models feature context windows up to 512,000 tokens and are trained on approximately 15 trillion tokens. The Granite 4.1 8B instruct model matches or outperforms the previous Granite 4.0 32B mixture-of-experts model while using a simpler dense architecture.
What is IBM Granite 4.1?
IBM Granite 4.1 is a family of dense, decoder-only language models designed for enterprise use cases. The family includes three sizes: 3B, 8B, and 30B parameters, each available in both base and instruct variants. All models use a dense transformer architecture rather than the mixture-of-experts approach used in the previous Granite 4.0 32B model.
The models are trained on approximately 15 trillion tokens across multiple phases, starting with broad pre-training and progressively moving toward higher-quality technical, scientific, and mathematical data with a focus on instruction following. Context windows extend up to 512,000 tokens, ensuring the models can process long documents without performance degradation on shorter-context tasks. IBM has released all Granite 4.1 models under the Apache 2.0 license, making them fully open source for commercial and research use.
Granite 4.1 benchmarks and technical specifications
The most notable performance claim for Granite 4.1 is that the 8B instruct model consistently matches or outperforms the Granite 4.0 32B mixture-of-experts model. This is significant because the 8B model uses roughly a quarter of the parameters with a simpler architecture, making it more practical for fine-tuning and deployment in resource-constrained environments.
The 30B model targets workloads that require stronger reasoning and generation quality, while the 3B model is optimized for edge deployment and latency-sensitive applications. IBM has also released Granite Speech 4.1 models alongside the language models, introducing multilingual speech recognition and translation capabilities designed for edge use cases.
Granite 4.1 availability and licensing
All Granite 4.1 models are released under the Apache 2.0 license, one of the most permissive open-source licenses available. This allows unrestricted commercial use, modification, and redistribution. The models are available on Hugging Face under the ibm-granite organization, on Ollama for local deployment, and through IBM’s enterprise AI platform.
The Apache 2.0 licensing distinguishes Granite 4.1 from many competing models that use more restrictive licenses. Compared to Mistral Medium 3.5’s modified MIT license or Meta’s Llama license, Apache 2.0 imposes fewer restrictions on derivative works and commercial applications. For enterprises evaluating open-source models, this licensing simplicity can be a deciding factor.
How does Granite 4.1 compare to other open models?
In the current landscape of open-weight models, Granite 4.1 competes primarily with Meta’s Llama family, Mistral’s models, and Alibaba’s Qwen series. The 8B model’s performance relative to the previous 32B MoE model suggests strong training efficiency, though direct benchmark comparisons with Llama 4 and Qwen 3 at equivalent parameter counts will determine its competitive positioning.
IBM’s focus with Granite has consistently been enterprise readiness: transparent training data documentation, strong instruction-following capabilities, and permissive licensing. The 512K context window places Granite 4.1 among the longest-context open models currently available, alongside Mistral Medium 3.5’s 256K window and recent Qwen models.
The Granite 4.1 family is available now on Hugging Face at huggingface.co/ibm-granite and through Ollama. IBM’s full technical blog post is available at research.ibm.com/blog/granite-4-1-ai-foundation-models.
