IBM has introduced its most advanced family of AI models to date, Granite 3.0 at the annual TechXchange event. According to the company, the third-generation Granite flagship language models has great performance, transparency, and safety, and on numerous academic and industry benchmarks, it can match or surpass comparably sized models from top model suppliers.
The Granite 3.0 models include general-purpose language AI models such as Granite 3.0 8B Instruct, Granite 3.0 2B Instruct, Granite 3.0 8B Base, and Granite 3.0 2B Base; Guardrails and Safety models such as Granite Guardian 3.0 8B, Granite Guardian 3.0 2B; and Mixture-of-Experts, including Granite 3.0 3B-A800M Instruct, Granite 3.0 1B-A400M Instruct, Granite 3.0 3B-A800M Base, Granite 3.0 1B-A400M Base.
IBM provides its Granite Mixture of Experts (MoE) Architecture models, Granite 3.0 1B-A400M and Granite 3.0 3B-A800M, as smaller, lighter models that could be used for CPU-based deployments as well as low latency applications, showcasing an outstanding balance between performance and inference cost.
Trained on Over 12 Trillion Tokens
The Granite 3.0 models are trained using a novel two-stage training method on over 12 trillion tokens of data from 116 different programming languages and 12 different natural languages. The results of thousands of experiments were used to optimize training parameters, data selection, and data quality.
IBM is also announcing an enhanced version of its pre-trained Granite Time Series models, which were first made available earlier this year. Outperforming 10 times larger models from Google, Alibaba, and other companies, these new models are trained on three times as much data and perform well on all three important time series benchmarks.
Additionally, the upgraded models offer more modeling flexibility by supporting rolling predictions and external factors.
Granite Guardian 3.0
IBM’s new family of Granite Guardian models enables application developers to put safety guardrails in place by examining user prompts and LLM answers for a range of risks. The Granite Guardian 3.0 8B and 2B models offer the most comprehensive set of risk and harm detection features available in the market right now.
Granite Guardian models also provide numerous unique RAG-specific checks such as context relevance, groundedness, and answer relevance.
According to IBM, the Granite Guardian 3.0 8B model outperforms all three generations of Meta’s Llama Guard models in terms of overall accuracy in harm detection, in comprehensive testing across 19 safety and RAG criteria.
Granite 3.0 Models Availability
Under the permissive Apache 2.0 license, the complete Granite 3.0 model suite and the updated time series models are available for download on HuggingFace. The instruct variations of the new Granite 3.0 8B and 2B language models as well as the Granite Guardian 3.0 8B and 2B models are now commercially available on IBM’s Watsonx platform.
Additionally, a number of the Granite 3.0 models will be accessible via HuggingFace integrations with Google Cloud’s Vertex AI Model Garden and as NVIDIA NIM microservices. A curated set of Granite 3.0 models can also be found on Replicate and Ollama.
Also Read:
- RBI Governor Raises Concerns Over AI in Financial Sector
- Entrepreneurial Innovation: What Impact Does AI Have on Entrepreneurs?
Stay Tuned to The Future Talk for more AI news and insights!