Gemma 4 Multimodal Fine-Tuner vs Lovable Desktop App

Gemma 4 Multimodal Fine-Tuner is an open-source toolkit that lets developers fine-tune Google's Gemma 4 and 3n models across all three modalities — text, images, and audio — using only Apple Silicon hardware. It runs natively on PyTorch with Metal Performance Shaders (MPS), bypassing the NVIDIA requirement that has historically blocked Mac users from serious local fine-tuning work. The toolkit handles the full training pipeline including dataset prep, LoRA adapters, and multi-modal data collation. It ships with working example notebooks, a validation suite, and clean abstractions that don't require deep familiarity with the underlying MPS stack. Apple Silicon's unified memory architecture actually helps here — large multimodal batches fit in memory that would otherwise require GPU VRAM splitting on CUDA setups. Posted to Hacker News on April 7 as a Show HN, it pulled 109 upvotes and 165 GitHub stars within hours. The timing is sharp: Gemma 4 just dropped days ago with new multimodal capabilities, and the community immediately wanted local fine-tuning. This fills that gap faster than Google's own tooling.

Lovable—the AI fullstack engineering platform with 35k+ followers and a 4.66/5 rating—launched its native desktop app today. The desktop version adds project tab organization for managing multiple AI-built apps simultaneously, and crucially: local Model Context Protocol (MCP) server support, letting Lovable agents connect to local services, databases, and tools running on your machine without routing through the cloud. Lovable's core product lets you build full-stack web applications by chatting with AI rather than writing code. It handles React frontends, Supabase backends, authentication, database schemas, and GitHub sync. The desktop app doesn't add new AI capabilities per se, but the local MCP integration is significant: it means Lovable agents can now talk to local Docker containers, local databases, or custom tools during the development process—something the browser version couldn't do. For the Lovable target audience—founders, indie hackers, and non-traditional developers building real products with AI—the desktop app signals the platform's maturation. Multi-tab project management alone reduces the friction of context-switching between different apps you're building. The local MCP support starts to make Lovable competitive with more developer-facing tools like Cursor for complex projects that need local environment access.