– New Granite 3.0 8B and 2B models, released under the permissive Apache 2.0 license, demonstrate strong performance across many academic and enterprise benchmarks, outperforming or matching similarly sized models .
– New Granite Guardian 3.0 model delivers IBM’s most comprehensive guardrail capabilities to advance safe and trusted AI
– New Granite 3.0 expert mixture model enables highly efficient inference and low latency, making it suitable for CPU-based deployments and edge computing
– New Granite Time Series model achieves state-of-the-art performance in zero/few shot predictions, outperforming larger models by a factor of 10.
– IBM announces next-generation watsonx Code Assistant powered by Granite for general-purpose coding. New tools for building and deploying AI applications and agents debut on watsonx.ai
– Granite announced to become the default model for Consulting Advantage, the AI-powered delivery platform used by IBM’s 160,000 consultants to deliver new solutions to customers faster
October 21, 2024
ARMONK, N.Y. , Oct. 21, 2024 /PRNewswire/ — Today at IBM’s (NYSE: IBM) annual TechXchange event, the company announces the release of Granite 3.0, its most advanced AI model family to date. I did. IBM’s 3rd generation Granite flagship language models outperform or match similarly sized models from leading model providers on many academic and industry benchmarks, delivering strong performance, transparency, and security. is shown.
Consistent with the company’s commitment to open source AI, Granite models are released under the permissive Apache 2.0 license and are unique in the combination of performance, flexibility, and autonomy they offer to enterprise clients and the community at large. It becomes.
IBM’s Granite 3.0 family includes:
General Purpose/Languages: Granite 3.0 8B Instructions, Granite 3.0 2B Instructions, Granite 3.0 8B Base, Granite 3.0 2B Base Guardrails and Safety: Granite Guardian 3.0 8B, Granite Guardian 3.0 2B Specialist Mix: Granite 3.0 3B-A800M Instructions, Granite 3.0 1B-A400M instruction, Granite 3.0 3B-A800M base, Granite 3.0 1B-A400M base
The new Granite 3.0 8B and 2B language models are designed to be enterprise AI “workhorse” models, delivering powerful performance for tasks such as search augmentation generation (RAG), classification, summarization, entity extraction, and tool usage. Masu. These compact and versatile models are fine-tuned for enterprise data and designed to seamlessly integrate across diverse business environments and workflows.
Many large-scale language models (LLMs) are trained on publicly available data, but a large portion of enterprise data remains untapped. By combining small-scale Granite models with enterprise data, specifically using InstructLab, an innovative tuning technology introduced by IBM and RedHat in May, enterprises can perform tasks comparable to large models at a fraction of the cost. IBM believes it can achieve unique performance (based on observed results) in some early proofs of concept, with costs ranging from 3x to 23x lower than larger Frontier models 1).
The Granite 3.0 release reaffirms IBM’s commitment to building transparency, security, and trust in AI products. The Granite 3.0 Technical Report and Responsible Use Guide provides a description of the datasets used to train these models, details of the filtering, cleansing, and curation steps applied, as well as models across key academic and corporate benchmarks. Provides comprehensive results of performance.
Importantly, IBM provides IP coverage for all Granite models on watsonx.ai, giving enterprise clients the ability to merge data with their models with more confidence.
Raising the bar: Granite 3.0 benchmark
The Granite 3.0 language model also shows promising results in terms of raw performance.
On standard academic benchmarks defined by Hugging Face’s OpenLLM Leaderboard, the overall performance of the Granite 3.0 8B Instruct model averaged out compared to the state-of-the-art performance of similarly sized open source models from Meta and Mistral. It’s excellent. In IBM’s state-of-the-art AttaQ safety benchmark, the Granite 3.0 8B Instruct model leads in all measured safety dimensions compared to the Meta and Mistral models. 2
Across core enterprise tasks such as RAG, tool usage, and cybersecurity domain tasks, the Granite 3.0 8B Instruct model performs better on average compared to similarly sized open source models from Mistral and Meta.3. performance.
Granite 3.0 models train 12 different natural languages using a new two-step training method that leverages the results of thousands of experiments designed to optimize data quality, data selection, and data selection. and trained on over 12 trillion tokens on data from 116 different programming languages. training parameters. By the end of the year, the 3.0 8B and 2B language models will include support for enhanced 128K context windows and multimodal document understanding.
Demonstrating an excellent balance between performance and inference cost, IBM offers its Granite Mixture of Experts (MoE) architecture models, Granite 3.0 1B-A400M and Granite 3.0 3B-A800M, as small and lightweight models that can also be deployed in low-latency applications. I am. As a CPU-based deployment.
IBM is also announcing an updated release of its pre-trained Granite Time Series model, the first version of which was released earlier this year. These new models are trained on 3x more data and perform better on all three major time series benchmarks, outperforming 10x larger models from Google, Alibaba, and more. The updated model also provides more modeling flexibility with support for external variables and rolling predictions. 4
Introducing Granite Guardian 3.0: Ushering in the next era of responsible AI
As part of this release, IBM is also introducing a new family of Granite Guardian models. This allows application developers to implement safety guardrails by checking user prompts and LLM responses for various risks. Granite Guardian 3.0 8B and 2B models offer the most comprehensive set of risk and hazard detection features currently available on the market.
In addition to aspects of harm such as social bias, hate, toxicity, profanity, violence, and jailbreak, these models incorporate a variety of unique RAG-specific aspects such as rationale, contextual relevance, and answer relevance. We also provide checks. In extensive testing across 19 safety and RAG benchmarks, the Granite Guardian 3.0 8B model was found to have higher overall harm detection accuracy on average than all three generations of Llama Guard models from Meta. It also showed an overall performance on average comparable to the specialized hallucination detection models WeCheck and MiniCheck in hallucination detection. 5
The Granite Guardian model is derived from the corresponding Granite language model, but can be used to implement guardrails alongside open or proprietary AI models.
Granite 3.0 model availability
The entire suite of Granite 3.0 models and updated time series models are available for download from HuggingFace under the permissive Apache 2.0 license. The new Granite 3.0 8B and 2B language model instruction variants and Granite Guardian 3.0 8B and 2B models are now commercially available on IBM’s watsonx platform. Some Granite 3.0 models will also be available as NVIDIA NIM microservices and through Google Cloud’s Vertex AI Model Garden and HuggingFace integration.
To provide developer choice and ease of use, and to support local edge deployments, a select set of Granite 3.0 models are also available in Ollama and Replicate.
The latest generation of Granite models expands on IBM’s robust open source catalog of powerful LLMs. IBM works with ecosystem partners such as AWS, Docker, Domo, and Qualcomm Technologies, Inc. via Qualcomm® AI Hub, Salesforce, SAP, and others to integrate various Granite models into their products. , and making Granite models available to our partners’ services. Our platform provides a wide range of choices to businesses around the world.
Agent Assistant: Enabling the future of enterprise AI
IBM is advancing enterprise AI through a variety of technologies, from models and assistants to the tools you need to tailor and deploy AI specifically for your company’s unique data and use cases. IBM is also paving the way for future AI agents that can self-direct, reflect, and perform complex tasks in dynamic business environments.
IBM continues to evolve its portfolio of AI assistant technologies. From WatsonX Orchestrate, which helps businesses build their own assistants through low-code tools and automation, to a wide set of pre-built assistants for specific tasks and domains such as customer service, human resources, and sales. ,marketing. Organizations around the world use WatsonX Assistant to answer everyday questions from customers and employees, modernize mainframe and legacy IT applications, and help students explore potential career paths. , we’ve helped build AI assistants for tasks like providing digital mortgage assistance to homebuyers.
Today, IBM also announced the next release of the next-generation watsonx Code Assistant, powered by the Granite code model. It has advanced application modernization features for the enterprise and provides general-purpose coding assistance across languages such as C, C++, Go, Java, and Python. The code capabilities of Java Applications.6 Granite are now also accessible through the Visual Studio Code extension, IBM Granite.Code.
IBM also plans to release new tools to help developers more efficiently build, customize, and deploy AI through watsonx.ai. This includes agent frameworks, integration with existing environments, and low-code automation for common use cases such as RAGs and agents. 7
IBM is focused on developing AI agent technology capable of increased autonomy, advanced reasoning, and multi-step problem solving. The initial release of the Granite 3.0 8B model features support for key agent features such as advanced reasoning, highly structured chat templates, and prompt styles for implementing tool usage workflows. IBM also plans to introduce new AI agent chat capabilities to IBM watsonx Orchestrate. This feature uses agent capabilities to orchestrate AI assistants, skills, and automation to help users improve productivity across their teams. 8 IBM plans to continue building agent capabilities across its portfolio. 2025 will include pre-built agents for specific domains and use cases.
Expanding AI-powered delivery platform to bring greater AI to IBM consultants
IBM is also announcing significant expansions to its AI-powered delivery platform, IBM Consulting Advantage. Multi-model platform includes methods such as AI agents, applications, and repeatable frameworks to enable 160,000 IBM consultants to deliver better and faster customer value at lower cost .
As part of the expansion, the Granite 3.0 language model will become the default model for Consulting Advantage. By leveraging Granite’s performance and efficiency, IBM Consulting can maximize the return on investment for IBM customers’ generative AI projects.
Another key part of the expansion is the introduction of IBM Consulting Advantage for Cloud Transformation and Management and IBM Consulting Advantage for Business Operations. Each includes domain-specific AI agents, applications, and methods that incorporate IBM best practices, allowing IBM consultants to help clients transform their cloud and AI with tasks such as code modernization and quality engineering. We can help you accelerate, transform and execute operations across domains such as finance and HR. and procurement.
To learn more about Granite and IBM’s AI for Business strategy, visit https://www.ibm.com/granite.
1 Cost calculations are API costs per million tokens for IBM watsonx for open models and openAI for GPT4 models (assuming a mix of 80% input/output and 20% output) for customer proof of concept. Based on the price.
2 IBM Research Technical Paper: Granite 3.0 Language Model
3 IBM Research Technical Paper: Granite 3.0 Language Model
4 Tiny Time Mixer: Fast pre-trained model to enhance zero/few shot predictions on multivariate time series
5 Publish evaluation results to Granite Guardian GitHub Repo
6 Scheduled to be available in Q4 2024
7 Scheduled to be available in Q4 2024
8 Scheduled to be available in Q1 2025
Media contact:
amy angelini
alangeli@us.ibm.com
Source IBM
