IBM is expanding its Granite AI family with models that combine reasoning, computer vision and predictive analytics.
The new Granite AI models feature reasoning, can analyze documents and predict trends. They are open-source and available through Hugging Face, IBM watson.ai and other platforms.
AI that thinks when needed
Granite 3.2 Instruct (8B and 2B) helps with summarization, problem solving and code generation. The model uses chain-of-thought reasoning, but IBM adds a twist: reasoning can be enabled or disabled at runtime. This saves computing power and makes the model more efficient than, say, DeepSeek R1, which always scans an entire reasoning space. According to IBM, this allows Granite 3.2 8B to compete with larger models such as Claude 3.5 Sonnet and GPT-4o on benchmarks such as AIME2024 and MATH500.
More than text: vision and predictions
In addition to text models, IBM is launching Granite Vision 3.2 2B, a multimodal model that analyzes documents and better handles fonts, graphics and infographics. For AI security, IBM is introducing Granite Guardian 3.2, which detects risks in prompts and indicates a confidence level rather than a simple yes or no.
Finally, IBM’s Granite Timeseries-TTM-R2.1 adds forecasting: this model analyzes trends up to two years ahead for applications in finance, supply chain and retail.
read also