Reasoning, vision, forecasting, and more in IBM Granite 3.2. The most recent version of the third generation of IBM Granite models, Granite 3.2, represents a crucial milestone in the series’ development beyond simple language models. Granite 3.2 adds a number of important new features to the Granite family, with experimental reasoning tools and the second official visual language model (VLM) taking Centre stage.
A number of enhancements to the effectiveness, efficiency, and adaptability of current offerings are also included in the update. IBM is pursuing state-of-the-art performance with an increasing number of parameters by prioritizing models that are practical and enterprise-ready.
The most recent Granite models are available under an Apache 2.0 license, as usual. Hugging Face currently offers all Granite models.
IBM Granite 3.2 Features
Granite 3.2 Instruct: Reasoning when you need it
In comparison to its 3.1 predecessors, the most recent versions of IBM’s flagship text-only large language models (LLMs), Granite 3.2 Instruct 8B and Granite 3.2 Instruct 2B, have been trained to provide improved reasoning capabilities. In keeping with IBM’s pragmatic approach to improving model performance, its application of reasoning considerably deviates from some industry trends.
IBM has integrated reasoning capabilities into the basic Instruct models, eliminating the need to release additional “reasoning models” and complicating development workflows. It is simple to turn on and off the model’s internal reasoning process, which guarantees that the right amount of computing power is used for the job.
IBM’s approach offers the advantages of reasoning while maintaining overall performance and safety, in contrast to conventional reasoning-driven approaches that enhance model performance on logical tasks (such math and coding) at the expense of other domains.
At IBM Research, there are several active investigations into reasoning-driven model evolution, of which instruct models are just one. Granite 3.2 8B Instruct can be calibrated to meet or surpass the mathematical reasoning performance of much bigger models, such as Anthropic’s Claude-3.5-Sonnet-1022 and OpenAI’s GPT-4o-0513, according to additional research on inference scaling methodologies.
Granite Guardian 3.2: Slimmer, safer, more specific
With reduced inference costs and memory utilization, Granite Guardian 3.2, the most recent version of IBM guardrail models created to identify hazards in prompts and responses, offers performance comparable to Guardian 3.1 counterparts at a faster rate.
The mixture of experts (MoE) base model, which only activates 800M of its 3B total parameter count at inference time, was refined to develop Granite Guardian 3.2 3B-A800M. With its launch, Granite Guardian now offers a particularly effective and affordable choice.
Granite Embedding: A new sparse embedding model
The most recent Granite Embedding model, Granite-Embedding-Sparse-30M-English, has a slightly modified architecture that allows it to learn sparse embeddings, in contrast to all previous Granite Embedding models (and, moreover, almost all embedding models in the current deep learning era) that learn dense embeddings.
Granite-Embedding-30M-Sparse strikes a compromise between scalability and efficiency across a range of resource and latency budgets, and is optimized for exact matches, keyword search, and English ranking. In order to accelerate the development cycle, it is made available through Granite Experiments, an IBM Research sandbox for testing open source concepts.
Granite Timeseries models: Now with daily and weekly forecasting
Hugging Face has received over 8 million downloads of IBM’s well-liked open source family of small Granite Time Series models, known as Tiny Time Mixers (TTMs). The latest model in the Granite Time Series portfolio, TTM-R2.1, covers daily and weekly forecasting horizons, whereas earlier TTM variations published within the TTM-R1 and TTM-R2 series allowed zero-shot and few-shot forecasting for minutely to hourly resolutions.
Getting started with Granite 3.2
Hugging Face offers all Granite 3.2 models under the permissive Apache 2.0 license. A few models can also be found on IBM Watsonx.ai and through platform partners like as Replicate, Ollama, and LM Studio, listed alphabetically. This article will be updated in the future to reflect Granite 3.2 models’ wider platform availability.
Granite documentation and the Granite Snack Cookbook on GitHub contain a variety of instructions and recipes for dealing with Granite models.
