Gemma 4 Multimodal Fine-Tuner vs Modo

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.

Modo is an open-source AI IDE built on the Void editor (a VS Code fork) that flips the script on how AI coding tools work. Instead of jumping straight to code generation, Modo forces a spec-first workflow: describe what you want, and the agent converts your prompt into structured requirements docs, design docs, and task breakdowns stored in a persistent `.modo/specs/` directory before writing a single line of code. The approach draws from the "vibe coding is bad actually" school of thought. Modo's steering files and agent hooks let developers set coding conventions, stack preferences, and project constraints that persist across sessions. Autopilot mode chains spec generation through implementation, while parallel chat lets you run multiple agent conversations simultaneously against the same codebase. Built by a solo developer and posted to Hacker News as a Show HN, Modo positions itself against Cursor, Windsurf, and Kiro. The bet: slowing down agents with structured planning up front produces fewer hallucinated architectures and rewrites. It's early — rough edges abound — but the spec-driven philosophy is increasingly mainstream as larger teams adopt AI coding tools.