VentureBeat
Follow
'Western Qwen': IBM wows with Granite 4 LLM launch and hybrid Mamba/Transformer architecture
IBM has launched Granite 4.0, a new family of open-source large language models designed for high performance and efficiency. This release marks IBM re-entering the competitive LLM landscape, particularly against Chinese models like Alibaba's Qwen. Granite 4.0 utilizes a novel hybrid architecture combining Transformer and Mamba designs. Transformers excel at context but are computationally expensive, while Mamba is more efficient for long sequences. This hybrid approach aims to leverage the strengths of both, significantly reducing GPU memory consumption by over 70%. The models are available under a permissive Apache 2.0 license, encouraging commercial use and modifications. Granite 4.0 demonstrates strong performance on benchmarks for instruction following and function calling. IBM emphasizes trust and safety, with Granite being the first open model family certified under ISO/IEC 42001. The models are trained on a vast 22-trillion-token corpus, including enterprise-relevant datasets. IBM plans further expansion with additional models for various enterprise needs. Granite 4.0 models are accessible through platforms like Hugging Face and IBM watsonx.ai, with broader partner support expected. This release positions IBM as a provider of enterprise-ready, cost-effective, and secure AI solutions.
