IBM announced the launch of its new generation open source large language model Granite3.1, aiming to lead the field of enterprise-level AI. Granite3.1 has many highlights, including extended context length to 128K, efficient embedding models, built-in hallucination detection capabilities, and significantly improved overall performance. According to IBM, its Granite8B Instruct model performs best among open source models of the same scale, surpassing competitors such as Meta's Llama3.1, Qwen2.5 and Google's Gemma2. The release of this new model follows the launch of Granite 3.0 in October, reflecting IBM's rapid iteration and continuous investment in the field of generative AI, with related business revenue reaching US$2 billion.
IBM officially released its new generation of open source large language model Granite3.1, striving to occupy a leading position in the field of enterprise-level AI. This series of models features 128K extended context length, embedding models, built-in hallucination detection capabilities, and significant performance improvements.
IBM claims that the Granite8B Instruct model performs best among open source competitors of the same size, including Meta's Llama3.1, Qwen2.5 and Google's Gemma2.
The release of the Granite3.1 model comes against the backdrop of IBM's rapid iteration of the Granite series. Granite3.0 was launched as early as October. IBM revealed that its business revenue related to generating AI has reached $2 billion. The core idea of the new version is to integrate more functionality into a smaller model so that it can be run more easily and cost-effectively for business users.
David Cox, vice president of IBM Research, said that the Granite model is widely used in IBM's internal products, consulting services and customer services, and is also released in open source form, so it needs to reach a high level in all aspects. Model performance evaluation relies not only on speed but also on efficiency, helping users save time when obtaining results.
In terms of context length, the improvement of Granite3.1 is particularly obvious, extending from the first version of 4K to 128K, which is particularly important for enterprise AI users, especially in terms of retrieval enhanced generation (RAG) and intelligent agent AI. The extended context length allows the model to process longer documents, logs, and conversations, allowing it to better understand and respond to complex queries.
IBM has also launched a series of embedding models to speed up the process of converting data into vectors. Among them, the query time of the Granite-Embedding-30M-English model is 0.16 seconds, which is faster than competitors' products. In order to achieve the performance improvement of Granite3.1, IBM has innovated in the multi-stage training process and the use of high-quality training data.
In terms of hallucination detection, the Granite3.1 model integrates hallucination protection into the model, which can self-detect and reduce false output. This built-in detection optimizes overall efficiency and reduces the number of inference calls.
Currently, the Granite3.1 model is open to enterprise users for free and is provided through IBM's Watsonx enterprise AI service. In the future, IBM plans to maintain a rapid pace of updates, and Granite 3.2 will launch multi-modal functionality in early 2025.
Official blog: https://www.ibm.com/new/announcements/ibm-granite-3-1-powerful-performance-long-context-and-more
Highlight:
IBM launched the Granite3.1 model, aiming to take a leading position in the open source large language model market.
The new model supports 128K context length, significantly improving processing capabilities and efficiency.
Illusion detection capabilities are integrated into the model, optimizing overall performance and accuracy.
All in all, the release of Granite3.1 marks another major progress for IBM in the field of open source large language models. Its powerful performance and rich functions will bring a more convenient and efficient AI experience to enterprise users. Future iterations are worth looking forward to.