IBM has launched the subsequent technology fashions in its Granite household: Granite 3.2 8B Instruct, Granite 3.2 2B Instruct, Granite Imaginative and prescient 3.2 2B, Granite-Timeseries-TTM-R2.1, Granite-Embedding-30M-Sparse, and new mannequin sizes for Granite Guardian 3.2.
Granite 3.2 8B Instruct and Granite 3.2 2B Instruct present chain of thought reasoning that may be toggled on and off. In response to IBM, chain of thought reasoning might be highly effective, however requires important computing energy that isn’t wanted for each process, which might result in pointless utilization.
The corporate took steps to mitigate this by permitting this characteristic to be simply turned off when it’s not wanted, and making use of Thought Desire Optimization (TPO)-based reinforcement studying, which permits it to attain better efficiency on advanced reasoning with out compromising efficiency elsewhere, the corporate defined.
“The discharge of Granite 3.2 marks solely the start of IBM’s explorations into reasoning capabilities for enterprise fashions. A lot of our ongoing analysis goals to make the most of the inherently longer, extra strong thought technique of Granite 3.2 for additional mannequin optimization,” IBM wrote in a weblog publish.
Granite Imaginative and prescient 3.2B is a brand new multimodal mannequin that was designed for doc understanding duties. In response to IBM, this mannequin matches or exceeds Llama 3.2 11B and Pixtral 12B on enterprise benchmarks together with DocVQA, ChartQA, AI2D, and OCRBench.
Granite-Timeseries-TTM-R2.1 extends the mannequin’s forecasting capabilities to now supply day by day and weekly predictions. Beforehand, it solely supported forecasting for minutes and hours.
Granite-Embedding-30M-Sparse is an evolution of the Granite Embedding fashions that now has the flexibility to be taught sparse embeddings, wherein their embedding dimension equals their vocabulary dimension, and might be considerably sooner than dense embeddings for shorter textual content passages.
The corporate can be releasing a 30% smaller Granite Guardian security mannequin, Granite Guardian 3.2 5B, that matches the efficiency of the earlier technology. Granite Guardian additionally has a brand new characteristic, verbalized confidence, offering a “extra nuanced danger evaluation that acknowledges ambiguity in security monitoring.”
IBM can be releasing Granite Guardian 3.2 3B-A800M, which was created by fine-tuning the corporate’s combination of specialists (MoE) base mannequin.
The entire new Granite 3.2 fashions can be found on Hugging Face underneath the Apache 2.0 license. Moreover, a few of the fashions are accessible by way of IBM watsonx.ai, Ollama, Replicate, and LM Studio.