IBM has announced the release of its latest family of AI models, promising a strong balance of power and cost efficiency for enterprises.
The company says its Granite 3.0 models offer smaller and more business-focused use cases, as opposed to many existing large-scale general-purpose language models that are already available.
More broadly, IBM also announced other AI improvements such as its latest watsonx Code Assistant, powered by Granite code models, which supports languages like C, C++, Go, Java and Python.
IBM Granite 3.0 models
IBM envisions the power of Granite 3.0 by combining what are essentially reasonably small models with enterprise data to unlock task-specific performance that rivals larger models, without the associated cost. The company declared that, across a range of testing, Granite 3.0 proved to be 3x-23x cheaper than large frontier models, which it failed to name.
The company did state that its 8B Instruct model – the larger of the two new 8B and 2B models – was able to keep up with similarly sized options from Meta and Mistral on benchmarks set out by Hugging Face’s OpenLLM Leaderboard.
At the same time, IBM also announced updates to its Granite Guardian 3.0 to allow developers to implement safety guardrails by checking user prompts and LLM responses. This includes checking for things like social bias, hate, toxicity, profanity, violence and jailbreaking. Another test confirmed that IBM’s solution gave Granite Guardian 3.0 8B improved accuracy on harm detection over Meta’s three generations of Llama Guard models.
All of the Granite 3.0 models are available for download on HuggingFace, with the 8B and 2B language models and the Granite Guardian 3.0 8B and 3B models also available fro commercial use through IBM’s watsonx.
Via Reuters
