IBM Granite drops compact speech model

This is a practical open-model release, not a hype-cycle moonshot: IBM is betting that smaller, cheaper speech systems with enterprise-friendly licensing will matter more than giant multimodal demos for real production use.

• –The headline improvement is efficiency: IBM says the model delivers faster inference and cuts parameter count in half versus Granite Speech 3.3 2B, which matters for self-hosted ASR pipelines.
• –Multilingual coverage across English, French, German, Spanish, Portuguese, and Japanese makes it more useful for global support, transcription, and translation workflows than English-only open models.
• –Keyword list biasing is a very enterprise feature; it directly targets names, acronyms, and jargon that usually break speech systems in business settings.
• –IBM ships usage examples for both Transformers and vLLM, which lowers adoption friction for teams already running open-source inference stacks.