IBM Releases Smaller, Faster & Open Source Granite AI Models
Global technology firms are expanding their enterprise AI offerings as corporate demand increases for business applications. IBM has launched its latest enterprise AI models under an open source licence.
The release marks a shift in the enterprise AI market, where most providers charge for access to their models. IBM’s Granite 3.0 family encompasses three model types designed for business use, which it says are smaller, faster and significantly cheaper – with IBM finding its new models are between three and 24 times cheaper than large frontier models in early proofs of concept.
Model architecture and capabilities
The Granite 3.0 family encompasses three model types released under the Apache 2.0 licence. The general purpose language models, available in 8B and 2B versions, handle text generation, classification and summarisation tasks. These models support Retrieval Augmented Generation (RAG), a technique that enables AI systems to incorporate external data sources.
The safety models, called Granite Guardian, establish ethical parameters across 19 safety benchmarks, including social bias, toxicity and hallucination detection. IBM reports these models demonstrate higher accuracy in harm detection compared to Meta's Llama Guard models.
The mixture-of-experts models enable deployment for low latency applications and CPU-based systems., while IBM says its new time series model claims performance surpassing models 10 times larger from competitors including Google and Alibaba.
Enterprise integration
IBM's strategy combines open source accessibility with commercial services. The company provides intellectual property indemnity for Granite models used through its watsonx platform, addressing enterprise concerns about data security.
The models integrate with IBM's watsonx Code Assistant, supporting languages including C, C++, Go, Java and Python. This integration extends to Enterprise Java Applications, with code capabilities accessible through a Visual Studio Code extension called IBM Granite.Code.
Business deployment
IBM trained these models on 12 trillion tokens spanning 12 natural languages and 116 programming languages. The company states the models outperform similar-sized open-source models from Meta and Mistral on standard academic benchmarks.
Ecosystem partners including Docker, Domo, Qualcomm Technologies, Salesforce and SAP will integrate Granite models into their offerings. The models will also be available through platforms including Ollama and Replicate for local and edge deployments.
IBM will make Granite 3.0 the default system in its Consulting Advantage platform, used by 160,000 consultants. The platform includes AI agents and applications for cloud transformation and business operations across finance, HR and procurement.
Kate Soule, Program Director of Data and Model Factory at IBM says: “Reflecting our focus on the balance between powerful and practical, the new IBM Granite 3.0 models deliver state-of-the-art performance relative to model size while maximising safety, speed and cost-efficiency for enterprise use cases.”
******
Make sure you check out the latest edition of AI Magazine and also sign up to our global conference series - Tech & AI LIVE 2024
******
AI Magazine is a BizClik brand