IBM has released its newest version of open-source Granite Large Language Models (LLMs), Granite 3.1, which marks a big step up in performance from earlier versions. These fresh models are fine-tuned to tackle business tasks and are set to beat existing models from rivals like OpenAI and Google in crucial areas.
Granite 3.1 has a context window of 128K tokens, which is a big step up from earlier versions. This larger window lets the models work with about 85,000 English words making it possible to do more thorough analysis and creation tasks. To put this in perspective, OpenAI’s ChatGPT 3, which kicked off the AI boom, could deal with 2,000 tokens.
IBM says its new Granite 8B Instruct model beats other models like Google’s Gemma 2, Meta’s Llama 3.1, and Qwen 2.5, according to the HuggingFace OpenLLM Leaderboard tests. The Granite 3.1 family has both dense models and Mixture of Experts (MoE) versions. These aim to help with different tasks such as tool-based use cases and retrieval-augmented generation (RAG) for jobs like coding, translating, and fixing bugs. The MoE models work well for on-device apps that need quick responses.
A standout upgrade in Granite 3.1 is its ability to handle images as input and produce text as output, enabling companies to bring visual content into their processes. The models have also grown their language skills now working with 12 languages such as German, Spanish, French, Japanese, and Arabic, which makes them a good fit for worldwide businesses.
Like earlier versions, Granite 3.1 models come under the open-source Apache 2.0 license. IBM keeps up its focus on openness by sharing the training datasets and giving full details on how they build and test the models. Also, to deal with worries about intellectual property, IBM offers unlimited protection against third-party IP claims that might come up when using its models.
The Granite 3.1 models are now available on IBM’s Watsonx platform and cloud services such as Google Vertex AI, Hugging Face, NVIDIA, and other AI platforms. IBM’s new models aim to speed up AI adoption in business settings giving companies strong, efficient, and reliable AI tools that can spark new ideas and tackle tricky problems at much lower costs than bigger models.
With Granite 3.1, IBM has made a big leap in offering open-source, business-grade AI solutions that are both powerful and budget-friendly making them a good choice for companies wanting to add AI to their work.
