What's New in Mellea 0.4.0 + Granite Libraries Release

The release of Granite libraries alongside Mellea 0.4.0 signals a critical inflection point for data engineering teams moving beyond third-party API dependencies. I've spent the last three years managing costs and latency around closed-model inference, and this shift toward production-grade open alternatives fundamentally changes our architecture conversations. We're now looking at self-hosted inference pipelines that let us control fine-tuning, maintain data privacy within our warehouses, and avoid vendor lock-in on proprietary endpoints. The operational reality is this: Granite libraries lower the barrier for embedding LLM capabilities directly into our dbt workflows and real-time feature pipelines without renegotiating contracts quarterly. However, this creates new ownership questions—data teams must now own model serving infrastructure and quantization strategies that were previously outsourced. My recommendation is to prototype these libraries in a bounded feature project first, ideally something with clear ROI like semantic search over your data catalog. This helps you understand the actual resource requirements before committing infrastructure budget.