IBM Granite

Granite 4.0 models (Micro, Tiny, Small) released in public beta, featuring hybrid Mamba/Transformer architecture for efficiency and low memory usage. Open-source under Apache 2.0, with full customizability and deployment flexibility. LangChain integration available for Replicate, enabling easy workflow orchestration. IBM announces bug bounty program (up to $100,000) and partnerships with EY, Lockheed Martin for enterprise testing. Roadmap includes larger and smaller models, as well as reasoning-focused variants by end of 2025.

Launch of Granite 4.0 with hybrid Mamba/Transformer architecture, reducing GPU memory consumption by over 70% and enabling deployment on consumer GPUs (e.g., NVIDIA RTX 3060). Models trained on 22T tokens from enterprise datasets (DataComp-LM, Wikipedia, curated subsets). Family includes Granite 4.0 Tiny (7B hybrid, 1B active parameters), Granite 4.0 Micro (3B dense hybrid), and Granite 4.0 Small (30B long-context instruct model). Post-training includes instruction-tuned and reasoning-focused 'Thinking' variants. Planned expansions: Granite 4.0 Medium (enterprise workloads) and Granite 4.0 Nano (edge deployments) by end of 2025.

Granite 3.2 introduces experimental reasoning capabilities and visual understanding (focus on document understanding). New Granite Guardian 3.2 models available on Hugging Face and watsonx.ai, with planned Ollama support. Embedding models (Granite-Embedding-30M-English, Granite-Embedding-107M-Multilingual) exceed rivals in inference speed. Open-source licensing and bug bounty program launched with HackerOne (up to $100,000 for vulnerabilities).

Launch of Granite 3.1 with long-range forecasting time series models (<10M parameters), optimized RAG and multimodal retrieval capabilities, and new Granite Guardian safety models with verbalized confidence for nuanced risk assessment. The 8B model achieves double-digit improvements in instruction-following benchmarks (ArenaHard, Alpaca Eval) and rivals larger models (Claude 3.5 Sonnet, GPT-4o) in math reasoning. Slimmed-down Granite Guardian models maintain performance with 30% size reduction.

Release of Granite-Instruct models, specifically fine-tuned for conversational AI applications. Introduction of a new 40B parameter model for edge deployment.

Granite models now support Retrieval Augmented Generation (RAG) natively within watsonx.ai. Improved handling of long-context inputs (up to 128k tokens).

Expanded language support to include Japanese, Korean, and Simplified Chinese. Reduced model latency by 15% through optimized inference.

Introduction of Granite 70B model. Added support for information extraction and question answering. Improved fine-tuning capabilities on watsonx.ai.

Improved performance on summarization tasks. Enhanced support for multilingual inputs (English, Spanish, French, German).

Initial release of the Granite family of models (7B, 13B, 34B parameters). Focus on text generation and instruction following. Available via watsonx.ai.